acceptodds
Withdrawn from ICLR 2027

DBFB: Bit-Aware Critical Weight Monitoring for Defending Bit-Flip Backdoor Attacks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Bit-flip backdoor attacks (BFBAs) constitute a sparse post-deployment threat in which an adversary flips only a few weight bits to activate targeted behavior while seeking to preserve clean accuracy. Existing bit-flip defenses are mainly designed for random faults, global accuracy degradation, or broadly sensitive parameters, and are therefore poorly aligned with the trigger-conditioned and highly localized nature of BFBAs. We propose DBFB, a selective defense that identifies and monitors BFBA-critical weights through a progressive neuron-to-weight-to-bit analysis. DBFB first locates attack-sensitive neurons by combining cross-class support with local boundary sensitivity and competition-direction concentration. It then maps these neurons to path-constrained candidate weights and evaluates their decision impact and structural propagation support. Finally, it re-ranks candidates using signed bit-induced changes and bounded estimates of relative margin erosion. The selected weights are stored with their clean reference values and checked before inference so that corrupted monitored weights can be detected and restored to their exact clean values. Across CNN and ViT architectures, four datasets, and three representative BFBAs, DBFB monitors only 5,000 weights, corresponding to 0.025% of model parameters on average. In the main evaluation, DBFB achieves 100% Hit Rate and 0% ASR for OneFlip on all four model–dataset pairs, and substantially suppresses the other evaluated attacks. The measured online check takes less than 0.1 ms, with approximately 39.1 KiB of index-and-reference storage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Bit-Flip Backdoor Attacks, Backdoor Defense, Deep Neural Network Security

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.40 over 5 reviews, above 9% of the others.
Reviewer_pa6MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DBFB, a post-deployment runtime defense designed to protect deep neural networks against trigger-conditioned bit-flip backdoor attacks (BFBAs) without requiring model retraining or architectural modifications.

Key observations is that parameters critical for general clean accuracy differ from those exploitable for targeted backdoor behavior, the authors propose a three-stage coarse-to-fine offline localization pipeline: (1) cross-class-constrained neuron selection, (2) path-aware weight ranking, and (3) bit-level fragility re-ranking based on bounded relative margin erosion.

At runtime, DBFB monitors the top 5,000 weights against an exact clean-reference and restores corrupted weights prior to inference, demonstrating 100% Hit Rate and 0% Attack Success Rate (ASR) against OneFlip and substantial suppression of ETBT and Deep-TROJ across CNN and ViT architectures on four small vision benchmarks.

Critical strengths
  1. The paper identifies an important distinction between generic clean-task parameter importance and targeted backdoor exploitability, motivating a training-free and fast defense with very compact overhead (~39.1 KiB).

  2. The proposed three-stage neuron-to-weight-to-bit localization hierarchy provides a structured clean-data surrogate model.

  3. Storing exact clean reference values for the selected critical weights enables simple swap at runtime.

Critical weaknesses
  1. By explicitly filtering out hidden-layer neurons with broad cross-class support and prioritizing concentrated single-class competition shifts to focus exclusively on targeted BFBAs, DBFB systematically excludes early-layer foundational feature extractors and leaves the model unprotected against sparse, data-free untargeted sign-bit and exponent-bit flip attacks that collapse global accuracy with just 1–2 bit flips (Maximal Brain Damage Without Data or Optimization: Disrupting Neural Networks via Sign-Bit Flips, 2025; Terminal Brain Damage: Exposing the Graceless Degradation in Deep Neural Networks Under Hardware Fault Attacks, 2019; Bit-Flip Attack: Crushing Neural Network with Progressive Bit Search, 2019). The first reference shown the exact models that the authors exemine.

  2. Operation is exclusive to full precision (FP32), leaving it unclear how DBFB generalizes to the standard 8-bit/4-bit fixed-point quantized models (INT8/INT4) and shared floating-point quantization scale factors targeted by the vast majority of bit-flip attacks (Bit-Flip Attack: Crushing Neural Network with Progressive Bit Search, 2019; TBT: Targeted Neural Network Attack with Bit Trojan, 2020; ProFlip: Targeted Trojan Attack with Progressive Bit Flips, 2021; TrojViT: Trojan Insertion in Vision Transformers, 2023; Backdoor Attacks on Neural Networks via One-Bit Flip, 2025; Your Scale Factors are My Weapon: Targeted Bit-Flip Attacks on Vision Transformers via Scale Factor Manipulation, 2025). Clearification: I do not say you need to showcase any of those attacks, but the quantized models.

  3. The empirical evaluation is limited to closed-set vision classification on small datasets, missing the common ImageNet-1K, where offline localization already requires 7,200 seconds on 200-class Tiny-ImageNet and Hit Rate drops to 1.88% under the fixed k=5000 budget. Moreover, Large Language Models or multimodal vision-language encoders (CLIP) where contemporary backdoor and bit-flip research is heavily concentrated and missing here.

  4. The experimental evaluation compares DBFB against only a single recent workshop baseline (WeightSentry): missing controlled top-k_w exact-restoration baselines and established runtime bit-flip defenses (Reference for defences: NeuroPots; Aegis; RADAR; HASHTAG;Aspis; ObfusBFA).

Questions for authors

Multi-Bit Attack Performance Gaps and Unreported ViT Multi-Bit Attack (Section 5.2 and Table 1): Please consider including a ViT-compatible progressive multi-bit BFBA (such as TrojViT or Flip-S) to provide a complete multi-bit evaluation on Vision Transformers, and discussing how KwK_{w} should scale on multi-bit attacks where partial restoration leaves noticeable clean accuracy degradation and residual ASR.

Reviewer_xTrAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes DBFB, a runtime defense against bit-flip backdoor attacks (BFBAs) that selects a fixed budget of scalar weights to monitor using only the clean model and clean data, through a three-stage ranking, and restores any monitored weight whose bit pattern deviates from a stored FP32 reference before each inference. It is evaluated on four model/dataset pairs against ETBT, Deep-TROJ and OneFlip with WeightSentry as the sole baseline, reporting 0% ASR on OneFlip in all settings and mixed results on the other two attacks.

Critical strengths

1. The problem is well motivated. Post-deployment bit-flip backdoors are sparse and trigger-conditioned, and existing runtime defenses are not designed for them. Selective monitoring with exact-reference restoration is a reasonable and cheap design, and the OneFlip results make the exact-restoration point concretely, since WeightSentry covers the same weight yet leaves ASR at 88–100%. The main evaluation spans CNNs and a ViT.

Critical weaknesses

1. Sections 4.1–4.3 present a sequence of mathematical constructions with little explanation of what each is meant to capture or why. Key quantities are undefined in the main text. For example, the relevance operator in Eq. (1) only appears as a formula in the appendix, and the activation and gate variables in Eqs. (6)–(8) are introduced without definition. The thresholds and aggregation rules throughout Section 4 are likewise stated without motivation. As a result, it is hard to see how the pieces fit together; even the fact that the stages form a sequential filter with an intermediate budget is only recoverable from the appendix. The burden of reconstructing the method's intent and checking its correctness is placed entirely on the reader. Section 4 needs to be rewritten rather than edited.

2. The relevance score, the sensitivity measures and the bit-level score all closely resemble existing attribution, gradient-sensitivity and attacker-side bit-search criteria, but none of this is acknowledged in the main text. If these constructions are the authors' own, each choice requires justification and a per-stage ablation. If they are adapted from prior work, as appears to be the case, the sources must be acknowledged and the novel elements clearly delineated. In either case, no argument is given for why these components should be combined in this order or why their composition should localise attack-relevant weights, which is essential to justify the method rather than report a pipeline that happened to work on the evaluated attacks.

3. After the support filter, every score ranks weights and bits by a first-order estimate of how much a perturbation erodes the clean-class margin on clean data. A successful BFBA by construction has almost no effect on clean behaviour. Instead, its effect is conditional on a trigger that activates a pathway that is quiet on clean inputs, where the clean-data gradient is close to zero. The quantity being ranked is therefore one the attacker explicitly minimises, and there is no reason to expect clean-data sensitivity to carry information about the trigger-conditioned behaviour the defense must anticipate. The paper itself states that attack criticality is not clean importance, then builds every score on clean-margin gradients. No direct evidence is provided that the scores predict attack locations beyond the Hit Rates themselves, which for the multi-bit attacks range from 1.9% to 100% across settings.

Questions for authors
  1. Please state Al,jA_{l,j} explicitly in the main text. Is al,ja_{l,j} taken before or after the nonlinearity? For a convolutional channel, is zl,jz_{l,j} a single scalar applied across all spatial positions? What precisely is a "hidden component along the classification-token readout path" in ViT-B-16?
  2. With the positive part applied to post-ReLU activations, many units in a layer may have Rl,j(x)=0R_{l,j}(x) = 0 exactly. Please report the per-layer distribution of Rl,j(x)R_{l,j}(x) (in particular the fraction that is exactly zero) so it can be determined whether this occurs, and explain how the states in Eq. (2) are assigned within that mass if it does.
  3. Please justify α=1/4\alpha = 1/4, the Nc/2N_c/2 and F1≥F2F_1 \ge F_2 rules in Eq. (3), ε=0.1\varepsilon = 0.1, t=3t = 3 and K=20,000K = 20{,}000, and report sensitivity to each. Figure 3 covers only KlK_l and KwK_w, on a single model/dataset pair.
  4. NeuroPots, BitShield, DeepNcode, SAVE and ALERT are all cited. Why is WeightSentry, a workshop paper, the only comparison? At least one further runtime or model-level baseline is needed, or a justification for each exclusion.
  5. Please report recovery accuracy (accuracy on triggered inputs against their true labels after the defense) alongside ACC and ASR. This is critical for understanding whether triggered inputs are correctly classified after the defense: it is possible for ASR to be low while recovery accuracy is also low.
  6. Please report mean and std over multiple seeds for Table 1, Figure 2 and Figure 3, and repeat the ablations on at least one pair other than ResNet-18/CIFAR-10.
  7. Each architecture in Table 1 is evaluated on a different dataset, so architecture and dataset effects are confounded. For example, it is not possible to tell whether the weak ETBT and Deep-TROJ results on VGG-16/GTSRB are due to the architecture or the dataset. Please evaluate at least one architecture across several datasets and one dataset across several architectures so that consistency can be assessed.
  8. WeightSentry achieves lower ASR than DBFB in four of the eleven entries in Table 1. This seems inconsistent with the abstract's claim that DBFB "substantially suppresses the other evaluated attacks."
  9. Please fix the broken reference in Algorithm 1 ("Eq. (??)") and the grammar in the third contribution bullet.
Reviewer_EQ9iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper addresses the vulnerability of deployed DNNs to post-deployment Bit Flip Backdoor Attacks (BFBAs), such as OneFlip, Deep-TROJ, and ETBT. These attacks try to flip a small number of weight bits in memory to inject a targeted backdoor. Authosr argue that the current defense models have misaligned objectives, where they focus on generic importance metrics, rather than identifying trigger-conditioned situations. Also, they identify the gap in inadequate runtime defenses. Range-clipping defenses (e.g., WeightSentry) do not restore exact clean values and can leave backdoors active even after clipping. The paper proposes an offline clean-data only profiling framework called DBFB, which hierarchically narrows candidates down across neuron, weight, bit levels. They use a monitoring set KwK_w for exact pre-inference restoration. Experimental results with multiple models show that DBFB outperforms WeightSentry baseline, achieving 0% residual ASR on single-bit attacks and they effectively suppress multi-bit attacks with minimal runtime latency.

Critical strengths
  1. The single-bit attack (OneFlip) defense results are promising, DBFB shows 100% hit rate and 0% residual ASR. Since DBFB exactly restores FP32 values, this neutralizes the backdoor. However WeightSentry's range-clipping leaves residual ASRs up to 100%.
  2. Restoring exact clean parameters offers a clear advantage, unlike the range-clipping defense baselin WeightSentry. Maintaining exact FP32 values for a small monitored set eliminates residual backdoor activation.
Critical weaknesses
  1. I find the presentation and overall structure of the paper quite disorganized and hard to understand. The figures/tables have little to no explanation in their caption. It makes it hard to follow what the paper actually tells.
  2. The major concerns for me are the lack of baselines, limited experiments and the actual contribution, detailed below:
  • Evaluation is limited to a single baseline (WeightSentry). There is no additional comparison between the mentioned model-level or runtime bit-flip defenses, such as ALERT (Wei et al., 2024), BitShield (Chen et al., 2025b), NeuroPots (Liu et al., 2023a).
  • Even though the model achieves 100% Hit Rate and 0% Defense ASR on single bit attacks (OneFlip), DBFB's localization capability drops significantly when facing multi-bit attacks: (e.g., Deep-TROJ hit-rates for ViT-B-1 or VGG-16). Restoring a small fraction of modified weights in a Deep-TROJ like attack offers no guarantee of backdoor suppression.
  • There is no quantized model evaluation. However, real-world post-deployment models subject to memory attacks (especially on edge devices) are almost universally quantized, either INT8 or FP8.
Questions for authors
  1. Under ETBT, DBFB performs worse than WeightSentry on ResNet-18/CIFAR-100 (15.79% vs. 0.94%) and VGG-16 (21.45% vs. 0.24%). Why was a fixed Kw=5000K_w = 5000 budget selected across all benchmarks, and does increasing KwK_w resolve this gap? The figure in supplementary (Fig. 3) does not show a budget sweep for the setups where ETBT struggles. It provides a KwK_w sweep only for ResNet-18/CIFAR-10, where ETBT was already at 0.99% residual ASR (Table 1).

  2. Does DBFB maintain its localization performance on quantized models (e.g., INT8), where weight shifts are strictly bounded in [−128,127], instead of the full-precision (FP32) models?

  3. I have concerns regarding the white-box threat model definition you've made. If an adaptive attacker knows DBFB monitors the top 5000 most fragile weights, what stops them from intentionally targeting weights ranked just outside this set to inject the backdoor while bypassing DBFB entirely?

Reviewer_J4qZOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DBFB, a runtime defense against bit-flip backdoor attacks (BFBAs) under a limited monitoring budget. It introduces a hierarchical neuron-to-weight-to-bit localization strategy to identify critical weights using clean data, followed by selective weight monitoring and exact-value restoration before inference. Experiments demonstrate promising defense performance across several models and attacks. However, insufficient methodological justification, limited baseline comparisons, and incomplete adaptive evaluations weaken the paper's contributions.

Critical strengths
  1. The paper addresses the important problem of defending deployed DNNs against BFBAs with limited monitoring resources. The proposed framework does not require model retraining or architectural modifications.

  2. DBFB effectively suppresses several evaluated attacks, achieving 0% ASR against OneFlip across all four settings while monitoring only 5,000 weights, with less than 0.1 ms of reported online checking time.

Critical weaknesses
  1. The paper compares DBFB primarily with WeightSentry, which uses range-based correction rather than exact-value restoration. Therefore, the improvements cannot be clearly attributed to the proposed weight-selection strategy. Moreover, WeightSentry outperforms DBFB in several cases, particularly ETBT on CIFAR-100 and GTSRB.

  2. Section 4.1 (Eqs. 4–9) prioritizes neurons based on cross-class support, margin sensitivity, and competition-direction concentration. However, the connection between these clean-data metrics and actual BFBA vulnerability remains insufficiently justified. Moreover, Figure 2 lacks individual ablations of the cross-class support constraint and competition-direction concentration, leaving their effectiveness unclear.

  3. The paper evaluates only three BFBA methods, with adaptive experiments restricted to OneFlip. Stronger evaluations against adaptive multi-bit attacks, such as ETBT and Deep-TROJ, as well as additional attack strategies such as TBT or ProFlip under compatible settings, are needed to establish the robustness and generalizability of DBFB.

  4. The hierarchical localization procedure involves multiple heuristics whose motivations and interactions are difficult to follow (Sections 4.1–4.3). In particular, the construction of path-conditioned parameter blocks (Eq. 11) and the combination of different sensitivity scores (Eqs. 9, 12, and 15) lack sufficient intuitive explanation and justification, making the method difficult to assess and reproduce.

Questions for authors
  1. Can the authors compare different weight-selection strategies under identical monitoring budgets and exact-value restoration mechanisms to isolate DBFB's actual contribution?

  2. Why should the cross-class support constraint and competition-direction concentration reliably identify BFBA-critical neurons? Can this assumption be validated empirically against actual attack-modified locations?

  3. Can the authors evaluate adaptive ETBT and Deep-TROJ attacks that explicitly avoid monitored weights, as well as additional unseen BFBA methods such as TBT or ProFlip?

  4. Can the authors clarify how path-conditioned blocks are constructed and justify the design choices and hyperparameters of the different scoring functions?

Reviewer_Hb3FOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper propose a DBFB method, which identifies and monitors the critical weights of bit-flip backdoor attack via the so called progressive neuron-to-weight-to-bit analysis. The proposed work is evaluated against three attacks, i.e., ETBT, Deep-TROJ and OneFlip, using four dataset including CIFAR-10, CIFAR-100, GTSRB and Tiny-ImageNet. The model used are small scale, ranging from ResNet-18, VGG-16 to ViT-B-16. Authors claim they can only monitor 5000 weights with 0.1ms online checking latency to suppress the aforementioned attacks.

Critical strengths

Authors claims their method can defend three popular bit-flip backdoor attacks with small checking latency.

Critical weaknesses
  1. First of all, the writing of manuscript should be significantly improved for better readability. The motivation of work is not properly presented and justified. The threat model of this work is not clearly presented.

  2. The original work of BFA attack [1] use the gradient-guided and bit-aware based method to progressive analyze the vulnerable parameter of the model. The innovativeness of this article is questionable. [1] Rakin, Adnan Siraj, Zhezhi He, and Deliang Fan. "Bit-flip attack: Crushing neural network with progressive bit search." 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019.

  3. The figure is poorly drawn. For example, text/labels in figure 1 are way too tiny to tell even after zoom-in. Figure 2 is highly like a AI-drawn figure with poor visualization.

  4. As this work propose a defense method, it does not properly compare with other competing defense method in a quantitative way. The reviewer cannot tell the advantages of proposed method.

  5. All the experiments are done on small scale model and easy tasks. Does the proposed method can scale up to large models? What is the overhead when scale to larger models?

  6. Moreover, the method only protect 5000 parameters. What if the attackers attack the weights not in the protected range? What if the proposed method is known to the attacker? Attacker can easily attack other weights of the model.

Questions for authors

check the weakness above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.