Modern computers store the weights of deployed neural networks in DRAM, and DRAM is physically fragile: as manufacturers shrink memory cells to fit more capacity onto a chip, neighboring rows disturb each other more easily, so a sufficiently persistent attacker can flip individual bits in a victim's memory (the Rowhammer effect). This is not a hypothetical concern. Researchers have shown that flipping as few as 13 bits out of 93 million in a ResNet-18 drops its ImageNet accuracy from 70% to 0.1% Bit-Flip Attack: Crushing Neural Network With Progressive Bit Search, that this can be done end-to-end on real DDR4 hardware in minutes using predictable operating-system memory allocation DeepHammer: Depleting the Intelligence of Deep Neural Networks through Targeted Chain of Bit Flips, and that it can even be triggered remotely without running any code on the victim machine Nethammer: Inducing Rowhammer Faults through Network Requests. The deployed defenses keep failing: hardware error correction is bypassed by multi-bit flips Another Flip in the Wall of Rowhammer Defenses, and software monitors either miss the attack or slow the memory system to a crawl SoftTRR: Protect Page Tables against Rowhammer Attacks using Software-only Target Row Refresh, QPRAC: Towards Secure and Practical PRAC-based Rowhammer Mitigation using Priority Queues.
The best existing runtime response, RADAR, detects corruption with cheap checksums but then repairs it by zeroing out the entire affected weight group — a permanent amputation RADAR: Run-time Adversarial Weight Attack Detection and Accuracy Recovery. This proposal argues that amputation is the wrong repair metaphor. Borrowing from three fields that have never been connected — network robustness simulation, which studies how graphs degrade under targeted node deletion, and neurorehabilitation, which studies when a transient intervention produces durable recovery rather than snap-back — we propose and test a plastic, dose-tuned repair: after detecting a bit flip, make small compensatory adjustments to surviving weights, sized by a personalized search, so the network re-routes its computation instead of losing it.
Core Claim
The claim has two parts. First, bit-flip damage in a neural network can be modeled as a targeted-attack process on a functional compute graph, where a corrupted high-magnitude weight acts like a deleted hub: the failure mode is not just a wrong number but a collapse in how information flows through the layers, and it has a predictable percolation-like threshold. Second, repair policy matters as much as detection: zero-out repair monotonically consumes the network's structural reserve until it crosses that threshold and collapses under repeated attack, whereas graduated compensatory repair — with the adjustment magnitude treated as a tunable "dose" — can produce recovery that is durable across attack rounds. We expect the dose to exhibit the inverted-U behavior documented in neurostimulation, where too little intervention does nothing and too much actively harms Personalized home based neurostimulation via AI optimization augments sustained attention.
This is deliberately different from the closest prior work. RADAR supplies runtime detection and a static, destructive recovery; we keep its detection idea and replace its recovery entirely. Aegis breaks attacker determinism with randomized early exits Aegis: Mitigating Targeted Bit-flip Attacks against Deep Neural Networks, but contamination of early backbone layers still propagates to every exit path — it raises attack cost without repairing anything. Stutz et al. harden weights at training time with clipping and random bit-error training On Mitigating Random and Adversarial Bit Errors, a static defense that adversarial search defeats with a handful of targeted flips DeepHammer. None of these works models the graph-level collapse dynamics, and none studies whether repair can persist under sustained attack. On the other side, the TIGER toolbox gives us mature machinery for targeted-deletion simulation and robustness metrics, but only on abstract infrastructure graphs, never on functional compute graphs Evaluating Graph Vulnerability and Robustness using TIGER. The therapeutic literature supplies the durability concept — interventions must cross a plasticity threshold to avoid elastic snap-back, and dose must be personalized — but has never touched artificial architectures Different doses of intermittent theta burst stimulation for upper limb motor dysfunction after stroke: a study protocol for a randomized controlled trial.
Research Program
Three components, each building on cited method work rather than reinventing it.
The first component builds the graph model. We represent a network's forward pass as a weighted directed graph whose nodes are channels (or neurons) and whose edge weights aggregate the magnitudes of the connecting weight matrices. A bit flip in a weight is then modeled as a distortion or severing of the incident edges, with severity given by the ratio of corrupted to original magnitude — the "outlier amplifier" mechanism that Stutz et al. and FT-ClipAct identify as the dominant failure mode, where a corrupted most-significant bit produces an explosive activation that blinds downstream subgraphs On Mitigating Random and Adversarial Bit Errors. We then run targeted-deletion simulations in the TIGER toolbox Evaluating Graph Vulnerability and Robustness using TIGER, using attack orderings from Progressive Bit Search as the deletion schedule and tracking global efficiency and giant-component size of the surviving activation graph. TIGER reports that average vertex betweenness can be accurately approximated using only about 10% of nodes, which makes adaptive recalculation — the expensive part of these simulations — tractable for our setting. The deliverable is an empirical percolation threshold: the cumulative fraction of effectively severed channels at which accuracy degrades discontinuously rather than gracefully, and a validated predictor of which weights are the hubs.
The second component is the repair mechanism itself. Upon checksum detection of a corrupted group (we reuse RADAR-style 2-bit signatures for detection only), we solve a bounded compensatory adjustment problem rather than zeroing:
$$\min_{\Delta}\;\mathcal{L}_{\text{val}}(w' + \Delta)\quad\text{s.t.}\quad \|\Delta\|_0 \le B,\;\;|\Delta_i| \le d$$
where $w'$ are the corrupted weights, $\Delta$ is a vector of adjustments restricted to surviving weights in the same and adjacent layers, $B$ is a small budget (tens to hundreds of weights), and $d$ is the dose bound — the maximum permitted adjustment per weight. The inner optimization is a few steps of gradient descent on a held-out validation batch, cheap enough for runtime use, and crucially it re-routes function instead of deleting capacity. DeepHammer's own observation that unintended extra flips sometimes helped the model recover is anecdotal evidence that compensatory perturbation can stabilize a corrupted network DeepHammer; this component makes that phenomenon deliberate and controlled.
The third component treats the dose $d$ (and budget $B$) as objects of closed-loop tuning, directly analogizing the personalized Bayesian optimization of Cohen Kadosh et al., which found that a fixed one-size-fits-all dose failed to beat placebo while individually tuned doses did ($p = 0.003$) Personalized home based neurostimulation via AI optimization augments sustained attention, and the dose-escalation logic of the iTBS trial protocol testing 600, 1,200, and 1,800 pulses to find the minimum threshold for stable motor recovery Different doses of intermittent theta burst stimulation. We will measure the recovery-vs-dose curve per model and attack type and test whether it is inverted-U shaped, whether durability requires a cumulative repair dose above some threshold (analogous to the 1,200-minute training threshold in stroke rehabilitation Effects of robotic gait training after stroke: A meta-analysis), and whether a schedule of spaced small repairs ("booster sessions") beats one large repair of equal total magnitude.
Baselines
The do-nothing reference point is the undefended network under attack. Concretely, Progressive Bit Search reports that flipping 13 of 93 million bits drops ResNet-18 ImageNet accuracy from 70% to 0.1% Bit-Flip Attack, and DeepHammer reaches random-guess accuracy with 2 to 24 bits on live DDR4 hardware DeepHammer. This baseline anchors how catastrophic inaction is, and its numbers also define the collapse point our durability metric must avoid.
The strongest existing competitor is RADAR RADAR: Run-time Adversarial Weight Attack Detection and Accuracy Recovery, because it is the only prior system that both detects and attempts recovery at runtime with near-zero overhead (under 1% inference time on gem5, about 5.6 KB of storage for ResNet-18, and detection of 9.6 out of 10 injected flips). We will run RADAR exactly as published on the same models, same attack sequences, and same hardware error profiles, giving it the same validation data and tuning budget our method receives. RADAR's reported recovery is from 0.18% to over 60% on ImageNet after 10 flips, and to over 80% on CIFAR-10 for ResNet-20 after 5 flips. Beating RADAR therefore means, first, recovering to within 2% of the clean baseline (roughly 68% on ImageNet where RADAR plateaus near 60%), and second — the decisive criterion — sustaining that recovery across repeated attack-repair rounds where RADAR's zeroing accumulates and eventually collapses the network. We also compare against Aegis Aegis: Mitigating Targeted Bit-flip Attacks against Deep Neural Networks, which reduced attack success rate from nearly 100% to below 10–20% by forcing roughly 35× more bit flips at under 2% benign accuracy cost; it is a prevention baseline rather than a recovery one, so we evaluate it on flips-to-collapse rather than post-attack recovery.
The key ablation isolates the repair policy from the detection machinery: identical checksum detection, identical attack schedule, with the recovery step swapped between zero-out and plastic adjustment. A second ablation fixes the dose $d$ at a single value versus closed-loop tuning, isolating the personalization ingredient. A third ablation applies the plastic repair without hub-prioritized initialization (random surviving weights instead of high-centrality ones), isolating the graph model's contribution.
Why not just…?
Why not just use ECC memory or on-die ECC? Because the literature shows it does not close the problem. Gruss et al. demonstrated hammering patterns that induce multi-bit flips exceeding SECDED correction Another Flip in the Wall of Rowhammer Defenses, and the commodity-OS survey confirms on-die ECC shifts but does not stop attacks SoK: Rowhammer on Commodity Operating Systems. ECC is also absent from edge and mobile devices where these models actually deploy. ECC raises the attack cost; it does not answer what happens when the residual attack succeeds, which is our question.
Why not just train robustly with weight clipping and random bit-error training? This is the cheapest credible defense, and Stutz et al. show it holds near-baseline accuracy at a 1% random bit error rate On Mitigating Random and Adversarial Bit Errors. But random noise training is the wrong threat model: targeted search defeats quantized networks in 2 to 24 flips DeepHammer, and when the target is compiled executable code rather than weights, static weight defenses provide no protection at all — an average of 1.4 flips suffices Compiled Models, Built-In Exploits: Uncovering Pervasive Bit-Flip Attack Surfaces in DNN Executables. We include RandBET-trained models as an additional baseline to measure how much residual damage plastic repair must handle on top of training-time hardening; the honest expectation is that the two are complementary, and our experiments will quantify that.
Why not just periodically reload the golden weights from secure storage? This looks sufficient and fails on timing. RADAR's own analysis notes that continuous integrity verification or periodic reloads violate real-time latency deadlines RADAR; between reloads, corrupted weights produce wrong classifications, and DeepHammer's memory massaging completes in under 100 seconds DeepHammer — well inside any realistic reload interval. Reload also assumes uncorrupted secure storage and an uncompromised load path. We will report reload latency on gem5 as a reference point so the comparison is quantitative rather than hand-waved.
Why not just use RADAR's zero-out, since it already recovers most of the accuracy? This is the real threat, and our answer is empirical, not rhetorical. RADAR recovers to roughly 60% on ImageNet against a 70% clean baseline, is explicitly weaker on non-MSB flips, and leaves permanent zeroed blocks RADAR. Our central durability experiment — repeated attack-repair cycles with the percolation threshold as the collapse criterion — is designed precisely to show where zero-out fails: each repair consumes capacity, so the network should reach its threshold after a predictable number of rounds. If plastic repair does not outperform zero-out on that curve, the proposal's core claim is wrong and the experiments will say so.
Why not just fine-tune the whole model after detecting an attack? Full fine-tuning is the upper bound on recovery quality and we will run it as such — but it is not a practical repair. It needs substantial compute, a full training or large validation set, and a pause in service, none of which edge inference deployments can afford mid-attack. The bounded adjustment in our formulation is best understood as a severely compute-limited fine-tune, and the dose study is what tells us how small the budget can be before recovery stops being durable. If the answer turns out to be "the budget must be nearly full fine-tuning," that is itself a publishable negative result bounding all lightweight repair schemes.
Proposed Experiments and Decisive Checks
Models and data follow the established benchmark set: ResNet-18 on ImageNet, ResNet-20 and VGG-16 on CIFAR-10, and CIFAR-100, STL-10, and Tiny-ImageNet for the Aegis comparison, with Google Speech Commands as a non-vision stress test, since all of these are the standard evaluation sets for the attacks and defenses we build on Bit-Flip Attack, DeepHammer, RADAR, Aegis. Attacks are Progressive Bit Search (white-box gradient-guided), DeepHammer (end-to-end with real DDR4 flip profiles and allocator massaging), and the compiled-model attack as a worst case where the corrupted object is executable code Compiled Models, Built-In Exploits. Runtime cost is measured on the gem5 architectural simulator, matching RADAR's evaluation methodology, with a target of staying under 5% inference overhead per repair event.
The first study validates the graph model: we sweep flips in PBS priority order, record accuracy alongside TIGER's global-efficiency and giant-component metrics on the functional graph, and test whether the discontinuity in accuracy coincides with the predicted percolation threshold. Success is a threshold predictor that generalizes across the three architectures with error small enough to rank defenses. The second study is the recovery comparison: after 5 and 10 flips (matching RADAR's reported settings), compare do-nothing, zero-out, plastic repair with fixed dose, and plastic repair with tuned dose, with full fine-tuning as the ceiling. The third study is durability: K successive attack-repair rounds, measuring accuracy, cumulative zeroed fraction, and the round at which each policy collapses; we predict zero-out collapses within a bounded number of rounds determined by the percolation threshold, while plastic repair does not amputate and therefore approaches the threshold only through attack damage itself. The fourth study is the dose-response curve: sweep $d$ and $B$ per model and attack, fit the inverted-U, and test the booster-schedule hypothesis that spaced small repairs beat one equal-magnitude repair. The compiled-model attack is the falsification stress test: if plastic weight repair cannot help when the corruption is in instruction code, the honest scope statement is that the method covers weight-plane attacks, and the graph model should predict this boundary from first principles — code-severing deletions sit upstream of every weight hub.
The decisive check for the whole proposal is a single curve: accuracy versus attack-repair round, plastic versus zero-out, on ResNet-18/ImageNet. If plastic repair sustains near-baseline accuracy across rounds where zero-out collapses, the durability thesis holds and the therapeutic analogy earned its place in systems security. If both collapse at the same round, then dose-tuned compensation adds complexity without durability, and the field should keep its checksums and its zeros.
References
- Bit-Flip Attack: Crushing Neural Network With Progressive Bit Search
- DeepHammer: Depleting the Intelligence of Deep Neural Networks through Targeted Chain of Bit Flips
- Nethammer: Inducing Rowhammer Faults through Network Requests
- Another Flip in the Wall of Rowhammer Defenses
- SoftTRR: Protect Page Tables against Rowhammer Attacks using Software-only Target Row Refresh
- QPRAC: Towards Secure and Practical PRAC-based Rowhammer Mitigation using Priority Queues
- RADAR: Run-time Adversarial Weight Attack Detection and Accuracy Recovery
- Personalized home based neurostimulation via AI optimization augments sustained attention
- Aegis: Mitigating Targeted Bit-flip Attacks against Deep Neural Networks
- Bit Error Robustness for Energy-Efficient DNN Accelerators
- Evaluating Graph Vulnerability and Robustness using TIGER
- Different doses of intermittent theta burst stimulation for upper limb motor dysfunction after stroke: a study protocol for a randomized controlled trial
- Effects of robotic gait training after stroke: A meta-analysis
- SoK: Rowhammer on Commodity Operating Systems
- Compiled Models, Built-In Exploits: Uncovering Pervasive Bit-Flip Attack Surfaces in DNN Executables