APT: Securing Against DRAM Read Disturbance via Adaptive Probabilistic In-DRAM Trackers

Defending DRAM against RowHammer means the chip must track which rows are being hammered and spend time refreshing their neighbors, and to keep both cheap, recent defenses sample rows with a fixed probability and borrow time through a DDR5 feature called Refresh Management — but this work from Huazhong University shows that fixed-probability sampling lets a clever attacker slip through, and that the Refresh Management commands leak a timing signal attackers can exploit; APT fixes both by adapting the sampling so no burst of activity escapes and by hiding the extra refreshes inside time the chip already spends refreshing, closing the evasion gap and the timing channel at once.

The argument in one glance

Problem

Defending DRAM against RowHammer costs space and time

RowHammer flips bits in a DRAM row by hammering its neighbors, and in-chip defenses must track which rows are being hammered (costing space) and refresh the victims (costing time). As read disturbance worsens, both costs grow, so defenses must be made cheap.
What people did

Sample rows probabilistically and borrow time via RFM

To cut space, recent defenses sample activations with a fixed probability using only a few tracker entries instead of counting every row. To find time, they use DDR5 Refresh Management (RFM) commands, which give the DRAM windows to run mitigations.
Why it fell short

Fixed sampling can be evaded, and RFM leaks timing

Fixed-probability sampling fails under dynamic activation patterns — the sampling misses let an attacker shape accesses to slip past the defense. And RFM commands cause measurable timing variations that become a side and covert channel leaking sensitive information. The defense is neither reliably secure nor free of new leaks.
This paper's idea

Adaptive sampling and hidden mitigations (APT)

APT (Adaptive Probabilistic In-DRAM Tracker) makes the mitigation probabilities across activations always sum to 1 — adaptively giving every activation an equal sampling chance based on real-time activation counts — so no burst escapes sampling. And because RFM is insecure, it hides mitigations in the unused slack of the normal refresh cycle while minimizing per-mitigation latency, avoiding the timing channel.

Why this problem is worth solving

RowHammer has grown from a curiosity into a standing threat that DRAM chips must actively defend against, and doing so cheaply is a real engineering constraint.

DRAM stores each bit as charge in a tiny capacitor, and to read a row the chip activates it, which slightly disturbs the physically adjacent rows. As cells have shrunk and packed closer together, this disturbance has grown strong enough that repeatedly activating one row — the aggressor — can leak enough charge from its neighbor rows, the victims, to flip their bits. This is the RowHammer family of read-disturbance attacks, and it lets an adversary corrupt data they were never allowed to write, which has been used to escalate privilege and break security. The standard defense is to detect a row being hammered and issue extra targeted refreshes to its victims before the bits can flip. Increasingly this defense is placed inside the DRAM chip itself, transparent to the memory controller, and such an in-DRAM defense needs two resources: space, to keep track of which rows are being activated often enough to be dangerous, and time, to actually perform the mitigating refreshes. As read disturbance worsens with each generation, both the number of rows to track and the number of mitigations to run grow, so keeping the defense low-cost is essential — a defense too expensive in area or performance will not be adopted.

The clever tricks the field uses to keep the defense cheap turn out to open two security holes, which is the crux of this paper.

To reduce the space cost, recent defenses do not count every row’s activations exactly, which would need a counter per row; instead they sample probabilistically, giving each activation some chance of being recorded in a small set of tracker entries, so a frequently-activated aggressor is likely to be caught with only a few entries. To find the time for mitigations, they use Refresh Management (RFM), a DDR5 feature where the memory controller, after enough activations, issues an RFM command that pauses normal traffic to give the DRAM a window to run its RowHammer mitigations. Both of these efficiency tricks introduce a flaw. First, if the sampling probability is fixed, it is only well-matched to one activation rate; when the real activation pattern is dynamic and bursty, a fixed chance produces row-sampling misses — windows where the aggressor simply is not sampled — and an attacker who understands the fixed probability can shape their accesses to fall into those gaps and evade mitigation. Second, an RFM command visibly stalls memory, so its occurrence causes measurable timing variations in other memory accesses; an attacker can observe these to infer activity (a side channel) or two colluding parties can trigger them deliberately to communicate covertly (a covert channel), so the defense itself leaks information. A truly secure defense must close both holes without giving up the low cost.

What people did before, and where it broke

Count every row’s activations exactly. Keep a precise per-row counter to detect any row crossing the hammer threshold. Where it broke. This requires far too much on-chip storage as densities rise, so it is not a low-cost in-DRAM solution.
Sample activations with a fixed probability. Use a few tracker entries and record each activation with a fixed chance, so aggressors are likely caught cheaply. Where it broke. Under dynamic, bursty activation patterns the fixed probability causes sampling misses, and an attacker can shape accesses to fall into those misses and evade the defense.
Borrow mitigation time with RFM commands. Use the DDR5 Refresh Management mechanism to give the DRAM windows to perform mitigations. Where it broke. RFM commands cause measurable timing variations that become a side and covert channel, so the defense leaks sensitive information.
Accept the efficiency-versus-security tradeoff. Treat cheap probabilistic-plus-RFM defenses as good enough despite the evasion and timing issues. Where it broke. It leaves both a coverage gap an attacker can exploit and a new information leak, so the defense is not actually secure.
The gap. No defense was simultaneously low-cost, un-evadable under dynamic activation, and free of the RFM timing channel. Providing that is what APT sets out to do.

How APT works

From fixed-probability, RFM-based defense to adaptive, hidden mitigation

↓ →
new
↓ →
new
↓ →
new
↓ →
new
↓ →
new

The idea

Keep the cheap probabilistic tracker, but remove its evadable gaps. APT holds onto the low-storage strategy of sampling activations into a few tracker entries rather than counting every row, because exact counting is too costly as read disturbance worsens; the problem was never sampling itself but the fixed probability that left exploitable holes. So APT keeps the small footprint while attacking the coverage weakness directly. Reducing the overhead of a RowHammer defense by tracking aggressors compactly rather than exhaustively, so protection stays affordable at scale, is the same goal the low-overhead RowHammer-mitigation work covered elsewhere on this site pursues, shrinking the cost of the defense so that securing memory against disturbance does not require prohibitive area or performance, while still catching the rows that matter.

Guarantee that no burst of activity escapes sampling. APT’s first key insight, per its abstract, is that the mitigation probabilities across activations should always sum to 1, and it achieves this by adaptively giving every activation an equal sampling chance based on the real-time activation count — so instead of a fixed per-activation probability that a shaped, bursty pattern can dodge, the sampling adjusts to the actual traffic and leaves no window uncovered. This turns an attacker’s ability to hide in the gaps of a fixed scheme into a guarantee that an attacked row will be captured. Closing an evasion path that an adversary would otherwise exploit by shaping memory accesses to slip past a defense’s blind spots is the same concern the row-coupling attack work covered elsewhere on this site raises from the offensive side, demonstrating how a RowHammer attacker can exploit exactly the mechanisms and gaps a defense fails to anticipate, which is what a robust, adaptive defense must foreclose.

Do not let the defense itself become a leak — hide the mitigations in existing slack. APT’s second key insight is that RFM is insecure because its commands cause observable timing variations, so rather than borrow time through RFM, the defense should exploit the unutilized slack within the normal refresh cycle (tRFC) to perform mitigations, while minimizing the latency of each one, so the extra work leaves no distinctive timing signature for an attacker to measure. This closes the side and covert channel that the RFM-based approach opened. Being careful that a security mechanism does not create a new memory-timing side channel through its own observable behavior is the same discipline the memory-timing side-channel work covered elsewhere on this site is built on, shaping how memory operations are timed so that an observer cannot infer secrets from the patterns a defense or scheduler leaves behind.

Following an attack against APT

Detecting and mitigating a hammered row without leaking timing

A conceptual trace based on the paper's abstract and standard RowHammer, RFM, and DDR5 background; specific mechanism and evaluation details were not available.

Step 1 of 5

An attacker hammers rows in a shaped pattern

An adversary repeatedly activates aggressor rows, deliberately varying the pattern to try to fall into the sampling gaps a fixed-probability defense would leave, hoping to flip a victim row's bits without being caught.

Dynamic, bursty activation is what defeats fixed sampling.

Why it works

Adaptive sampling has no fixed gap to exploit. Because APT sets sampling chances from the real-time activation count so probabilities sum to 1 → every window is covered rather than left to a fixed, guessable chance → so an attacker cannot shape accesses to fall into an unsampled gap.
A small tracker keeps the space cost low. Because sampling records aggressors into a few entries instead of counting every row → the storage stays modest even as densities and hammer counts rise → so the defense remains a low-cost in-DRAM solution.
Hiding mitigation in refresh slack removes the timing signature. Because APT runs mitigations within the unused slack of the normal refresh cycle rather than via RFM → the extra work does not produce the distinctive stalls RFM causes → so the side and covert channel is closed.
Minimizing per-mitigation latency preserves both goals. Because each mitigation is kept short → more mitigations fit within the available refresh slack without visible delay → so security is achieved without reintroducing a timing cost or an RFM dependency.
Attacking cost and security together avoids the usual tradeoff. Because the two insights target the evasion gap and the timing leak separately → APT can be cheap and secure at once → so it escapes the efficiency-versus-security compromise prior probabilistic-plus-RFM defenses made.

What it costs (trade-offs and scope)

This explainer is based on the abstract, not the full paper. The full text was not openly available at the time of writing, so the exact adaptive-sampling algorithm, how the probabilities are computed and made to sum to 1, precisely how mitigations fit within tRFC slack, the security analysis, and the evaluation numbers are not reflected here beyond what the abstract states, combined with standard background.
Fitting mitigations in refresh slack has limits. Exploiting the unutilized slack within the refresh cycle assumes there is enough slack for the needed mitigations; under extreme attack rates or very aggressive future hammer thresholds, the available slack could become a constraint the full design must manage.
Adaptive sampling needs real-time counts. Basing sampling chances on real-time activation counts requires the defense to maintain and use those counts cheaply; the benefit depends on that adaptation logic being low-cost, which is part of the design the abstract asserts but does not detail.
It targets the probabilistic-plus-RFM design point. APT is positioned against recent low-storage probabilistic defenses that use RFM; its advantages are relative to that class, and comparisons to exact-counter defenses or other mitigation styles are part of the full evaluation not available here.
Security claims need the full analysis. The abstract asserts that summing probabilities to 1 prevents sampling misses and that avoiding RFM closes the timing channel, but the strength of these guarantees against a determined adversary rests on the paper’s detailed security analysis, which is not available here.
It is an in-DRAM hardware mechanism. APT is a defense built into the DRAM, so deploying it requires DRAM support for its tracking and refresh-slack mitigation; it is not a software or memory-controller-only change.

The evaluation

Low-cost in-DRAM RowHammer defenses compared

Property	Exact per-row counters	Fixed-probability sampling + RFM	APT (adaptive + refresh slack)
Storage cost
Robust under dynamic activation
Evadable by shaped accesses	No
Mitigation timing channel	Depends on mechanism
Both low-cost and secure	Secure but costly

Low-cost in-DRAM RowHammer defenses, as framed by the APT abstract (the full paper was not openly available, so this reflects the abstract's claims and standard background rather than the paper's detailed evaluation). Exact per-row activation counters are robust but need too much on-chip storage as read disturbance worsens. Recent low-storage defenses instead sample activations with a fixed probability into a few tracker entries, and borrow mitigation time using DDR5 Refresh Management (RFM) commands — cheap, but with two security holes: under dynamic, bursty activation the fixed probability causes sampling misses that an attacker can shape their accesses to exploit and evade the defense, and RFM commands cause measurable timing variations that form a side and covert channel leaking sensitive information. APT (Adaptive Probabilistic In-DRAM Tracker) keeps the small tracker but fixes both. It ensures the mitigation probabilities across activations always sum to 1, adaptively giving every activation an equal sampling chance based on the real-time activation count, so no window is left unsampled and an attacked row is reliably captured. And because RFM is insecure, it exploits the unutilized slack within the normal refresh cycle (tRFC) to run mitigations while minimizing per-mitigation latency, so the mitigations leave no distinctive timing signature. The stated goal is a defense that is simultaneously low-cost, un-evadable under dynamic activation, and free of the RFM timing channel.Underlined values explain themselves when clicked.

Where this stops working

What to keep in mind

This explainer is based on the paper’s abstract, because the full text was not openly accessible when it was written, so read it as a faithful account of the abstract’s claims plus standard, well-established background on RowHammer, in-DRAM defenses, probabilistic sampling, DDR5 Refresh Management, and memory-timing side channels — not as a report of the paper’s detailed mechanism or measurements. With that caveat, the contribution is sharp. DRAM read disturbance keeps worsening, so chips carry in-DRAM defenses that cost space to track hammered rows and time to refresh victims, and the field has kept those costs low with two tricks: probabilistic sampling into a few tracker entries, and borrowing mitigation time via DDR5 Refresh Management commands. The paper’s insight is that each trick opens a security hole. A fixed sampling probability is only tuned for one activation rate, so under dynamic, bursty patterns it suffers sampling misses that a knowledgeable attacker can steer their accesses into, evading mitigation; and RFM commands visibly stall memory, so they become a side and covert channel leaking information — the defense undermines itself. APT closes both: it makes the sampling probabilities across activations sum to 1, adapting to the real-time activation count so every window is covered and no shaped pattern can hide, and it abandons RFM in favor of hiding mitigations in the unused slack of the normal refresh cycle, minimizing per-mitigation latency so nothing distinctive is observable. The result is meant to be low-cost, un-evadable, and leak-free at once. What is not available here is exactly how the adaptive probabilities are computed, how much refresh slack the mitigations need and how they are scheduled into it, the formal security analysis backing the two guarantees, and the quantitative overheads and comparisons — all the specifics needed to judge and reproduce it. The durable idea, clear even from the abstract, is that efficiency shortcuts in a security mechanism can themselves become vulnerabilities: a cheaper tracker that samples with a fixed probability creates an evasion gap, and a convenient way to schedule mitigations creates a timing channel, so a defense is only as secure as its own side effects, and the fix is to make coverage adaptive and to hide the defense’s activity within work the system already does.

Background

Why RowHammer defense keeps escalating on both sides

It helps to see APT within an ongoing arms race between DRAM read-disturbance attacks and the defenses meant to stop them, a race that intensifies with every denser memory generation because fewer activations are needed to flip a bit. Each new defense must be cheaper than the last to be adoptable, yet each efficiency shortcut invites a new attack, and each new attack demands a more careful defense. The comprehensive DRAM-disturbance-defense work covered elsewhere on this site reflects the same frontier, wrestling with how to protect memory against the full range of disturbance patterns without the mitigation cost dominating — exactly the low-cost-yet-secure tension APT navigates, here with the twist that the mitigation machinery itself, through RFM timing, had become an attack surface. The common thread is that DRAM security is not a one-shot fix but a moving target: defenders trade space, time, and secrecy against attackers who probe every gap and every observable side effect, including the defense’s own. APT’s specific lessons — that fixed-probability sampling leaves an evasion gap best closed by adapting to real activity, and that scheduling mitigations through an observable command leaks timing best avoided by hiding in existing refresh slack — are instances of a broader discipline that will keep recurring: as the hardware gets more fragile and the defenses more elaborate, the defenses must be scrutinized as adversarially as the attacks they prevent.

If you remember only three things

1. DRAM read disturbance — the RowHammer family, where repeatedly activating one row flips bits in its neighbors — keeps worsening as chips densify, so DRAM carries transparent in-chip defenses that cost space (to track hammered aggressor rows) and time (to refresh the victim rows). To keep these cheap, recent defenses use probabilistic row sampling with only a few tracker entries (catching aggressors by chance rather than counting every row) and borrow mitigation time using DDR5 Refresh Management (RFM) commands, which pause traffic to give the DRAM windows to run mitigations. But this probabilistic-plus-RFM approach has two security holes. First, a fixed sampling probability is tuned for one activation rate, so under dynamic, bursty patterns it produces row-sampling misses, and an attacker who knows the fixed probability can shape their accesses to fall into those gaps and evade mitigation. Second, RFM commands cause measurable timing variations in memory access, turning the defense itself into a side and covert channel that leaks sensitive information.
2. APT (Adaptive Probabilistic In-DRAM Tracker), per its abstract, fixes both holes while keeping the small, low-storage tracker. Its first insight is that the mitigation probabilities across activations should always sum to 1: APT adaptively allots every activation an equal sampling chance based on the real-time activation count, so no window is left unsampled and an attacked row is reliably captured, replacing the fixed, evadable probability with adaptive, gap-free coverage. Its second insight is that RFM is insecure because its commands are observable, so instead of borrowing mitigation time through RFM, APT exploits the unutilized slack within the normal refresh cycle (the tRFC timing) to perform mitigations, while minimizing the latency of each one, so the extra work leaves no distinctive timing signature for an attacker to measure. Together these aim for a defense that is simultaneously low-cost, un-evadable under dynamic activation, and free of the RFM timing channel.
3. Important caveat: this explainer is based on the abstract because the full paper was not openly available, so the exact adaptive-sampling algorithm, how mitigations are scheduled into refresh slack, the formal security analysis, and the quantitative overheads and comparisons are not reflected here. The durable idea, clear even from the abstract, is that the efficiency shortcuts in a security mechanism can themselves become vulnerabilities: a cheaper tracker that samples with a fixed probability creates an evasion gap, and a convenient way to schedule mitigations (RFM) creates a timing side and covert channel — so a defense is only as secure as its own side effects. The fix APT exemplifies is to make coverage adaptive to real activity so nothing can hide in a gap, and to hide the defense's activity within work the system already does (normal refresh) so it leaks no timing, closing the evasion and the channel at the same time.

Check yourself

Check your understanding

1. What two resources does an in-DRAM RowHammer defense need, and how do recent defenses try to keep each cheap?
2. Why do fixed-probability sampling and RFM each create a security vulnerability?
3. What are APT's two key insights for being both low-cost and secure?
4. What should you keep in mind about the evidence behind this explainer?

Sources

The paper is “APT: Securing Against DRAM Read Disturbance via Adaptive Probabilistic In-DRAM Trackers” by Runjin Wu, Meng Zhang, You Zhou, Changsheng Xie, and Fei Wu, published at ASPLOS 2026, DOI 10.1145/3779212.3790126. The authors are affiliated with Huazhong University of Science and Technology.

This explainer is based on the paper’s abstract (Creative Commons Attribution / CC-BY, as indicated by the publisher’s open-access metadata). The full text was not openly accessible at the time of writing — the ACM Digital Library copy was gated behind a bot check, there was no arXiv or author preprint available, and the OpenAlex, Semantic Scholar, and Unpaywall records pointed only to the gated version — so this page reflects the abstract’s stated claims (that worsening DRAM read disturbance makes transparent in-DRAM defenses need space to track aggressor rows and time to perform mitigations; that recent works use probabilistic row-sampling with few entries to reduce storage and Refresh Management, or RFM, commands introduced in DDR5 to find time; that probabilistic defenses with RFM face two challenges, namely fixed-probability sampling under dynamic activation causing row-sampling misses that allow attacks to evade mitigation, and RFM causing timing variations exploitable for side and covert channels; and that APT, the Adaptive Probabilistic In-DRAM Tracker, builds on two insights — ensuring mitigation probabilities across activations sum to 1 by adaptively allotting equal sampling chances based on real-time activation counts to prevent sampling misses, and exploiting the unutilized slack within the refresh cycle, tRFC, while minimizing per-mitigation latency because RFM is insecure) together with standard, well-established background on RowHammer and DRAM read disturbance, in-DRAM mitigation via targeted refresh, probabilistic activation sampling, DDR5 Refresh Management, and memory-timing side and covert channels, provided to make the abstract’s contribution understandable. The specific adaptive-sampling algorithm, the scheduling of mitigations within refresh slack, the formal security analysis, and the quantitative overheads and comparisons are not reflected here beyond what the abstract states; the flow and step-through sections are conceptual readings of the abstract plus standard background rather than details drawn from the full paper. This page is marked as a draft for that reason and should be confirmed against the full paper when it becomes available.