acceptodds
Under review as a conference paper at ICLR 2027

Towards Verifying Neural Networks Against Multi-Parameter Bit-Flip Perturbations

Abstract

Hardware faults can flip bits in the stored weights of a quantized neural network, potentially compromising its predictions. While such faults typically affect multiple parameters simultaneously, existing verifiers are limited to single-parameter perturbations due to the combinatorial explosion of possible flip locations in large networks. We present mBFV (-BitFlip Verifier), an efficient verification framework that proves robustness against simultaneous bit flips across multiple parameters without explicitly enumerating these combinations. mBFV achieves this via a novel multi-parameter bound propagation technique that directly aggregates the worst-case contributions. To further tighten these bounds, mBFV employs a branch-and-bound mechanism over perturbation locations, partitioning the potential flips to smaller groups of neurons. Evaluated on 625 instances, mBFV successfully verifies 293, significantly outperforming a prior single-parameter verifier (38 verified instances) and an exact mixed-integer linear programming baseline (0 verified instances). Notably, while these baselines are restricted to single-parameter flips on small networks (up to 13k parameters), mBFV scales to verify networks with up to 1.15M parameters against up to four simultaneous parameter flips.

Primary area:
interpretability and explainable AI
Keywords:
dnn verification, bit flip attack

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FvAJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the robustness verification of quantized neural networks (QNNs) when bit flips occur simultaneously across multiple parameters. The authors propose mBFV, which constructs per-parameter contribution bounds within a CROWN-style bound propagation framework. By employing a top-m aggregation strategy to account for perturbations in up to mm parameters, the method avoids the need to explicitly enumerate all possible combinations of perturbed parameters. To enhance the precision of the relaxation, the authors further introduce a Branch-and-Bound (BaB) approach tailored to the locations of the perturbations. Experimental evaluations—covering seven fully connected architectures across the ACAS Xu, MNIST, and CIFAR-10 datasets—demonstrate the effectiveness of the proposed verification method through comparisons with baselines and an ablation study of the BaB component.

Critical strengths
  • The research problem is meaningful, and the methodology is driven by a clear computational motivation. Verifying multi-parameter bit flips is more challenging than verifying single-parameter perturbations. This paper incorporates a cardinality constraint on parameter perturbations into bound propagation and combines it with a perturbation-location-based BaB approach, offering a viable alternative to the direct enumeration of parameter combinations.

  • The BaB approach yields clear empirical gains. In the reported experiments, the number of verified instances increased from 184 to 293 when BaB was applied. This ablation study demonstrates that partitioning based on perturbation locations helps mitigate the excessive looseness of a uniform relaxation, serving as a key piece of evidence for the effectiveness of the proposed method.

Critical weaknesses

Section 3 lacks a clear narrative, which severely compromises the readability of the core methodology. The authors present a dense sequence of formulas, often accompanied by only one or two sentences of explanation. It is difficult for the reader to discern: what specific problem is being addressed at this stage? Why is this particular formula introduced? What is its function, and how does it relate to the subsequent step? Such explanations should be provided by the authors; instead, the reader is left to piece them together from the formulas themselves.

My primary concern is not the omission of derivations for certain formulas—it is acceptable to skip routine computational steps, provided that the underlying logic of the method is clearly articulated. This is especially critical regarding the paper’s core contribution: how the margin lower bound is derived when multiple parameters are perturbed simultaneously, and why this bound is reliable. Rather than developing a coherent explanation around these points, Section 3 relies on a barrage of formulas to carry the entire explanatory burden. This makes the core contribution difficult to grasp and hinders the reader’s ability to assess its soundness.

In my view, Section 3 requires a structural reorganization, rather than merely the addition of a few missing derivation steps. The authors should first outline the overall approach, then clarify the purpose of each step and how these steps interconnect. Formulas should serve to support the narrative, not replace it. Incorporating a small, running example—such as a simple network—throughout the section would also be beneficial; however, the essential task remains to clearly explain the methodology. This current style—a mechanical enumeration of formulas—is my principal reservation regarding the paper.

Questions for authors
  1. Narrative and Soundness in Section 3. The current presentation in Section 3 makes it difficult to assess the soundness of the proposed method. While the authors provide numerous formulas, the purpose of each step—and how these steps collectively yield a reliable lower bound on the margin—remains unclear. Please reorganize this section to clearly articulate the core logic and the basis for its soundness. I will not adjust my current rating until this issue is addressed.

  2. Practical Application of Certificates. How are the certificates produced in this paper intended to be used? Are they meant for verifying specific input regions prior to deployment, or for verifying new inputs individually at runtime? If the goal is to ensure that all inputs classified as "A" are robust against parameter faults, is it necessary to verify each input individually, or can the verified regions be designed to cover these inputs? How would such coverage be implemented in a practical system, and what is the associated computational overhead?

  3. Model Scale and Experimental Contribution. Why do the experiments focus primarily on fully connected networks of the current scale, rather than evaluating larger or more complex model architectures? Is this a limitation inherent to the method itself, or is it due to excessive computational cost? If the choice of models is based solely on the fact that "prior work used them," I find this justification insufficient.

6 comments
AuthorsOpenReview ↗

Response to Reviewer FvAJ - Revised Narrative in Sec. 3

We thank the reviewer for the thorough critique and constructive suggestions. We have substantially revised Sec. 3 to address your concerns regarding narrative and readability, and we address each question below.

Q1 & W1. Narrative and soundness in Sec. 3

We restructured Sec. 3 to prioritize conceptual clarity over formula density. Instead of having derivations in isolation, the section is now organized around the two core challenges of multi-parameter bit flip verification:

The first challenge is propagating bounds through non-linear layers when parameters are perturbed. Sec. 3.1 shows that backward linear bound propagation separates unperturbed propagation from additive parameter perturbations. This isolates the parameter effect into sensitivity coefficients g‾p≤0\underline g_p\leq 0, which are evaluated in constant time from bounding hyperplanes and perturbation intervals.

The second challenge is selecting the worst-case combination of mm perturbed parameters out of KK without enumeration. Sec. 3.2 resolves this by relaxing binary parameter indicators to the continuous domain Cm=s∈[0,1]K:∑psp≤mC_m=\\{s\in [0,1]^K:\sum_p s_p\leq m\\}. We prove that this continuous relaxation incurs zero relaxation gap, as its minimum is always attained at a binary vector (Prop. 3.1). This yields an exact closed-form solution via top⁡m(g‾)\operatorname{top}_m(\underline g) in O(K)O(K) time.

Concretization and soundness then follow directly in Sec. 3.3. Because input bounds and parameter perturbations decouple additively in Eq. (10), minimizing them independently in Eq. (12) yields a provably sound lower bound.

To ground the methodology, Sec. 3.2 now walks through a concrete 4-parameter example (g=(−5.5,−3.2,−1.7,0)g=(-5.5,-3.2,-1.7,0) at m=2m=2) demonstrating why continuous relaxation matches the integer optimum. Furthermore, Apdx. A provides a complete numeric trace across a two-neuron network, showing intermediate bounds and sensitivity values at every step.

Q2. Practical application of certificates

How are the certificates produced in this paper intended to be used? Are they meant for verifying specific input regions prior to deployment, or for verifying new inputs individually at runtime?

Safety-critical applications have safety specifications defining the required behavior over specified input ranges. Even millions of successful tests cannot establish the absence of violations. Our verifier takes these specifications and checks them entirely prior to deployment. A certificate guarantees the required behavior for every input in the specified region under the allowed parameter faults.

If the goal is to ensure that all inputs classified as "A" are robust against parameter faults, is it necessary to verify each input individually, or can the verified regions be designed to cover these inputs?

Yes, our sound bound propagation method covers every possible input within the region given by the specification, e.g., neighbor of a specific input in class “A”. The bounds can be conservative (loose), but they contain all possible behaviors within the region, and mBFV verifies all input in the defined input range all at once.

How would such coverage be implemented in a practical system, and what is the associated computational overhead?

As noted above, the input ranges in the safety specification are verified entirely offline. Once a range is certified, every input satisfying its conditions preserves the required output property under the allowed faults. All verification cost is incurred offline, with zero additional computation during inference.

AuthorsOpenReview ↗

Response to Reviewer FvAJ - Experiment

Q3. Model scale and experimental contribution

The fully connected models in the submission were chosen to provide direct comparisons with existing state-of-the-art tools (BFAVerifier and MILP) on standard verification benchmarks. The scale is not an inherent limitation of our method: while existing baselines time out on networks with more than 13k parameters even for single bit flips (m=1m=1), mBFV scales to 1.15M parameters at m≤2m\leq 2 and 575k parameters at m=4m=4.

To demonstrate applicability to more complex architectures, the revised paper evaluates a convolutional model: CIFAR-10 CNN-A (215k parameters, two convolutional and two fully connected layers) from SDP-CROWN [a]. Verified instances out of 15 selected inputs (60 s timeout):

radius 0radius 1/255
m=1m=1, mBFV w/o BaB1514
m=1m=1, mBFV1515
m=2m=2, mBFV w/o BaB00
m=2m=2, mBFV1510
m=4m=4, mBFV00

At radius 1/255, each method is evaluated on instances it verified at radius 0. While BFAVerifier and MILP verify zero instances (timing out on all 15 inputs), mBFV verifies all 15 inputs at m≤2m\leq 2 (radius 0) and 10 inputs at radius 1/255. Crucially, branch and bound is essential: without BaB, zero instances verify at m=2m=2 (Sec. 5.1, Tab. 1).

References

[a] Chiu et al. SDP-CROWN: Efficient Bound Propagation for Neural Network Verification with Tightness of Semidefinite Programming. ICML 2025.

Thanks for your answer and clarification. I have one more question.

Have you considered comparing the bounds obtained under the original model and the bit-reversed model in same setting? Have you considered comparing the bounds obtained under the original model and the bit-reversed model?

I'm sure you'll come to some interesting conclusions.

Reviewer_2JQLOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper combines the bound propagation with branching over fault locations to efficiently verify quantized NNs when multiple faults in parameters. The main idea has three main components: represent possible maximum and minimum parameter changes, propagate bounds with a global parameter budget, branch over the nodes when subdivide groups cannot certify the network.

Critical strengths

Strong points:

  1. The extension of the single-parameter verification to multi-parameter problem is useful. Supporting simultaneous faults across several parameters addresses a meaningful gap beyond single-parameter certificates. The specification also combines parameter faults with input uncertainty
  2. A branching strategy has clear empirical value. Branching increases the total certificates from 184 to 293. Its contribution is especially substantial at m=4m=4, where certificates increase from 21 to 74 on page 8.
  3. On page 12, substantial attention to correctness arguments. The manuscript includes coverage proofs, a worked example explaining relaxation loss, and reported Lean mechanizations of several mathematical components. These make the theoretical claims easier to scrutinize.
  4. GPU batching and memory-saving aggregation help make larger networks tractable. The appendix also correctly emphasizes that convolutional weight sharing must be preserved when counting perturbed parameters.
Critical weaknesses

Weak points:

  1. On page 6, the main branching pseudocode omits an important initialization case. Algorithm 2 initializes the queue using sets of exactly mm layers. If m>Lm>L, this queue is empty, and the written algorithm could return “verified” after the initial bound fails. Appendix A.2 supplies a different initialization using BRANCH(all layers) when L≤mL\le m, but this correction is absent from the main algorithm. This is a concrete specification inconsistency; it does not establish that the implementation has the same defect. It matters directly to the shallow MNIST network evaluated at m=4m=4.
  2. The headline scaling claim needs qualification. On Page 1-2, the wording about networks with 1.15M parameters and four simultaneous faults can suggest a joint achievement. Table 1 instead shows zero four-fault certificates for the largest network.
  3. The complexity claim is easy to overinterpret. The search bound grows as νm\nu^m for fixed mm, and counts bound propagations rather than their total computational cost. Moreover, parameter enumeration is also polynomial in KK when mm is fixed. The meaningful improvement is the search representation and practical pruning, not a general transition from exponential to polynomial complexity.
Questions for authors
  1. On page 7, how much of the observed success depends on fine-tuning and clipping? Could you report certificate rates, clean accuracy, and fault-search results across several clipping values and training seeds?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2JQL - Soften Claims & Clarify about Fine-tuning.

We thank the reviewer for the careful reading and constructive feedback. We address each point below.

W1

Thank you for catching the omission and it was only in the pseudocode.

In the revised paper, we updated Alg. 2 and Sec. 4.1 with a conditional check: when L>mL>m, the queue initializes with all subsets of mm layers; when L≤mL\leq m, it initializes by branching directly on the root group covering all layers. Experimental results are unaffected.

W2

We agree that the original phrasing could suggest four simultaneous flips were certified on the 1.15M model. In the revised Abstract and Intro, we explicitly separate the two claims: mBFV verifies models up to 1.15M parameters at m≤2m\leq 2, while certifying four simultaneous flips on models up to 575k parameters.

W3

We agree with the reviewer's comment: for fixed mm, parameter enumeration is also polynomial (O(Km)O(K^m)).

In the revised paper (Sec. 1, Sec. 4.2, and Apdx. B.3), we removed claims of polynomial complexity of BaB and reframed the bound strictly as characterizing the worst-case search space is O(νm)O(\nu^m) subproblems.

As the reviewer noted, the meaningful advantage is the search representation and practical pruning. Branching over neurons practically reduces the search space base from parameters KK to neurons ν\nu with ν≪K\nu \ll K and shrinks candidate pairs from 3.6×10103.6\times 10^{10} to 541k subproblems on MNIST 2×2562\times 256 at m=2m=2).

Q1

The number of verified instances depends directly on fine-tuning and weight clipping because both govern intrinsic network robustness [a]. Because verifiers can only certify models that are genuinely robust, all methods (BFAVerifier, exact MILP, and mBFV) evaluate the exact same fine-tuned and clipped weights on identical inputs, ensuring a fair comparison.

Without fine-tuning and clipping, weights span wide ranges where single bit flips alter predictions, so verifiers certify zero instances on standard pretrained networks. On MNIST FC 2x256 (98.1% accuracy), mBFV verifies 0/10 inputs at m=1 and baselines time out. Fine-tuning with clipping (c=0.1, 96.2% accuracy) yields verifiable models, allowing mBFV to certify 15/15 inputs at m<=2 and 14/15 at m=4 at radius 0, whereas baselines time out without finding faults.

Weight projection ([-c, c]), an established defense against bit errors in quantized networks [a], trades clean accuracy for bit-flip robustness: bit flips perturb parameters by at most about 2c2c. On MNIST FC 2x256, evaluation on 3 inputs at radius 0 confirms that both c=0.1 and c=0.05 verify 3/3 inputs at m <=4; however, halving c to 0.05 drops accuracy by 13 points (83.2%).

Training MNIST 5x100 across three random seeds with c=0.05 (4-bit quantization, radius 0.02, and 10 test inputs per seed) demonstrates that these verification rates remain most stable across training independent runs:

trainingaccuracyverified (m=1)verified (m=2)verified (m=3)
hinge loss + clipping (seeds 0 / 1 / 2)92% / 93% / 86%9/10 / 8/10 / 8/108/10 / 8/10 / 8/104/10 / 2/10 / 4/10
standard, no clipping (seed 0)99%1/100/10–

Across three seeds, mBFV consistently verifies 8–9/10 inputs for three fine-tuned and clipped models, whereas unclipped standard training yields at most one verified instance. BaB gains are equally consistent. Without branching, mBFV certifies 8/10 at m=1 but drops to zero at m >= 2 across all seeds.

References

[a] Stutz et al. Bit Error Robustness for Energy-Efficient DNN Accelerators. MLSys 2021.

Reviewer_2JQLOpenReview ↗

Final comments

Thanks the authors for providing clarification answers. The response partially resolves the concern. I would retain a weak-accept recommendation, but these conditions do not justify a strong acceptance. The evidence is reasonably convincing that mBFV improves certification on the evaluated models; it remains limited evidence of broad practical effectiveness.

We thank you for your constructive feedback. We will include the clarification in the revision.

Reviewer_tU4dOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed a bound propagation framework for verifying whether a quantized neural network remains robust against ℓp\ell_p-norm attack when several stored network parameters suffer bit flips. The authors also tighten their method by branching on the possible locations of bit flips, and show that the number of branching needed is polynomial in the number of neurons. The simulation result demonstrates the proposed algorithm is more scalable than SOTA such as BFAV and MILP. It also demonstrates the proposed branching strategy is highly effective in tightening their method, at least in the verification for MNIST networks.

Critical strengths
  1. The proposed method is compatible with existing bound propagation methods, making it possible to scale to large networks.
  2. The branching strategy proposed in the paper has an upper bound on the number of branching needed to perform, which is polynomial in the number of neurons.
  3. This paper is really well written and organized.
Critical weaknesses
  1. The simulation needs to include a border range of models. So far, only feedforward networks are evaluated. The paper should at least include some convolutional networks.
Questions for authors
  1. Recently, SDP-CROWN integrated linear bound derived from SDP relaxation into α\alpha-CROWN, leading to tighter ℓ2\ell_2 verification. Can the proposed method be also applied to extend SDP-CROWN?

  2. Can you clarify the definition of (m,n)(m,n)-perturbation vector? Is it equivalent to ρ=(αt,Pt)∣t≤m,∣Pt∣≤n\rho = \\{(\alpha_t,P_t)\mid t\leq m, |P_t|\leq n \\} where Pt⊆1,…,nP_t\subseteq\\{1,\ldots,n\\}.

  3. You use g‾p\underline g_p (and g‾p\overline g_p) in both equation 9 and 10, are they supposed to be computed separately? It seems like in equation 11, you are computing the worst case A‾D\underline AD and worst case A‾d\underline A d separately and then add them together. Can you instead compute the combined worst case A‾D+A‾d\underline AD+\underline A d, wouldn’t that make your problem tighter? For instance, g‾p=min⁡D,zA‾D+A‾d\underline g_p=\min_{D,z}\underline AD+\underline A d.

3 comments
AuthorsOpenReview ↗

Response to Reviewer tU4d - CNN-A Result

We thank the reviewer for the helpful feedback and positive assessment. We address each point below.

W1. Convolutional networks

We agree, and the revised evaluation includes a convolutional model. We evaluate CIFAR-10 CNN-A (215k parameters) from SDP-CROWN [a], a standard verification benchmark. mBFV handles convolutions natively by treating each kernel entry and bias as an individual stored parameter subject to bit flips.

Verifying CNNs under parameter faults is harder than fully connected models: our default seed finds no verified inputs at m=2m=2, so we select 15 inputs with highest bounds from mBFV w/o BaB at m=2m=2 and radius 0 (reported in Sec. 5.1 and Apdx. D, outside Tab. 1 totals).

Verified instances out of 15 inputs (60 s timeout):

radius 0radius 1/255
m=1m=1, mBFV w/o BaB1514
m=1m=1, mBFV1515
m=2m=2, mBFV w/o BaB00
m=2m=2, mBFV1510
m=4m=4, mBFV00

At radius 1/255, each method is evaluated only on instances it verified at radius 0. While BFAVerifier and exact MILP verify zero instances, mBFV verifies all 15 inputs at m≤2m\leq 2. Crucially, branch and bound is essential: without BaB, zero instances verify at m=2m=2.

Q1. Extension to SDP-CROWN

Short answer is Yes.

We note that SDP-CROWN [a] tightens input-space ℓ2\ell_2 verification while mBFV handles parameter-space bit flips. During back-substitution, we could use SDP-CROWN's coupled ReLU bounds while collecting our per-parameter terms g‾p\underline g_p, with intermediate domains covering both perturbations.

Concretization step would then add our exact top-mm selection of parameter contributions to the ℓ2\ell_2 input bound, giving g(α)⊤x^+h(g(α),λ)−ρ∥g(α)∥2+top⁡m(g‾)g(\alpha)^\top \hat x+h(g(\alpha),\lambda)-\rho\Vert g(\alpha)\Vert_2+\operatorname{top}_m(\underline g). Finally, our BaB could apply this combined bound within each subproblem.

Q2. Def. of (m,n)-perturbation vector

We thank the reviewer for suggesting this concise definition of (m,n)(m,n)-perturbation vectors. We adopted this concise notation in Def. 2.1: ρ=(pr,Pr)∣r≤m,;∣Pr∣≤n,Pr⊆1,…,Q\rho = \\{(p_r, P_r) \mid r \le m, \\; |P_r| \le n\\}, \quad P_r \subseteq \\{1, \dots, Q\\} with distinct parameters prp_r. Here mm bounds perturbed parameters, prp_r indexes the parameter, PrP_r specifies flipped bits, and nn limits flips per parameter.

Q3. Separate vs. Joint Evaluation of Sensitivity Terms

Yes, we compute separate contributions for individual weights and biases.

For context, the submission's Eqs. (9), (10), and (11) correspond to Eqs. (8), (9), and (10) in the revision:

  • Weight p=Wjk(i)p=W_{jk}^{(i)}: g‾p=min⁡D∈[ℓp,up],,z∈[z‾i−1,k,z‾i−1,k]A‾jDz\underline g_p=\min_{D\in[\ell_p,u_p],\\,z\in[\underline z_{i-1,k},\overline z_{i-1,k}]}\underline A_jDz.
  • Bias p=bj(i)p=b_j^{(i)}: g‾p=min⁡d∈[ℓp,up]A‾jd\underline g_p=\min_{d\in[\ell_p,u_p]}\underline A_jd.
  • Combined bound: z^ℓ,t≥A‾x+b‾+∑pspg‾p\hat z_{\ell,t}\ge\underline A x+\underline b+\sum_p s_p\underline g_p, where s∈0,1Ks\in \\{0,1\\}^K and ∑psp≤m\sum_p s_p\leq m.

Each sps_p indicates whether one stored parameter is perturbed, allowing the worst mm contributions from any flipped weight or bias to be selected globally via topm⁡(g‾)\operatorname{top_m}(\underline g). Joint evaluation is possible, but merging a weight and bias into one contribution counted once can loosen the bound by allowing two flips for one selection. Keeping one contribution per parameter preserves the global limit of mm perturbed parameters and makes top-mm selection exact for the decomposed contributions..

Evaluating each parameter independently assigns an individual coefficient g‾p\underline g_p, allowing Eq. (10) to select the worst mm parameters globally in closed form via top⁡m(g‾)\operatorname{top}_m(\underline g) in O(K)O(K) time.

References

[a] Chiu et al. SDP-CROWN: Efficient Bound Propagation for Neural Network Verification with Tightness of Semidefinite Programming. ICML 2025.

Thank you for the clarification and additional evaluation. I have no further questions or concerns.

We thank you for your feedback and comments. We'll revise the paper to include clarifications and new results.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Gentle reminder

Dear reviewers, please let us know what you think about our rebuttal. We believe we have addressed all questions but would be happy to engage in more discussions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.