- Rating
- 3(+1)
- Confidence
- 2(+1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates the adversarial robustness theory of rate-coded SNNs. By refining the Lipschitz bounds for Bernoulli-smoothed classifiers, the study improves the dependence of the certified robustness radius on the number of time steps from exponential decay to inverse-linear decay, for a fixed class-probability margin, while also extending the certified robustness guarantee to signed rate encoding. Methodologically, the paper proposes SEA-PGD, an iterative sparse attack targeting the encoded spike sequences. Experiments on CIFAR-10, CIFAR-100, and SVHN demonstrate that SEA-PGD is more potent than the original SEA attack. The results further indicate that robustness guarantees in the input space do not inherently cover attacks in the encoding space, though adversarial training can enhance defense performance.
Critical strengths
The exponential decay of the certified radius with respect to time steps is improved to linear decay.
Critical weaknesses
Equation (14) is written as . Taking , we get , and after the attack , violating the specified $ \{-1,0,1\} $ constraint. It may just be a typo, but the implementation must be checked to determine if it affects the experiment.
Questions for authors
In Table 1, the clean accuracy of a certain signed model in SVHN is 84.35%, but it increases to 93.05% after the SEA attack. While an increase in accuracy after an attack is not entirely impossible, the magnitude of this increase requires explanation.
9 comments
Thank you for the clarification. I have two remaining concerns:
- In Appendix A.1 (p. 14, lines 707–708), the triangle inequality is written as an equality. In Appendix A.2 (p. 16, line 843), differentiability at zero is incorrectly assumed; signed encoding can yield an expectation proportional to . Moreover, coupling the encodings gives
suggesting that the Lipschitz bound for can be reduced from to , doubling the certified radius. Please verify these points and revise the proofs and constants accordingly.
- The shortcut explanation is plausible but remains unverified. To adequately address the increase from 84.35% to 93.05%, please: (i) report correct-to-incorrect and incorrect-to-correct prediction transitions, and recompute robust accuracy while retaining zero perturbation as a valid candidate; (ii) test the label-leaking hypothesis, for example by replacing true labels with predicted labels when generating attacks; and (iii) provide paired/repeated evaluations to control encoding randomness, together with implementation checks confirming that the corrected Eq. (14) satisfies the spike-value and sparsity constraints.
We are grateful to the reviewer for the detailed reading of the manuscript.
-
(i) We have corrected the typo related to the triangle inequality on p. 14, lines 707–708.
(ii) It is correctly highlighted that the function is non-differentiable at ; we should have been more careful. Fortunately, this does not break the proof of Lipschitzness. To simplify the exposition, consider the case with , where for signed rate encoding. We have,
\begin{align} g(x) = \mathbb{E}_{z \sim sBer(x)}[f(z)]= f(1) x + f(0) (1-x) \quad \text{if } x\in [0,1]\ \text{and, } g(x) =f(-1) (-x) + f(0) (1+x) \quad \text{if } x \in [-1,0] \end{align} So that,
\begin{align} \lvert g'(x) \rvert = \lvert f(1)- f(0) \rvert \leq 1 \quad \text{if, } x\in (0,1] \text{and,} \quad \lvert g'(x) \rvert = \lvert f(0) -f(-1) \rvert \leq 1 \quad \text{if } x \in [-1,0) \end{align}
We have piecewise on the intervals and , except at where it is non-differentiable. However, is continuous at , with . By the Mean Value Theorem, a continuous function that is differentiable on finitely many subintervals and whose derivative is bounded in absolute value by on each subinterval is globally -Lipschitz.
This ensures that the function is -Lipschitz, for . We would extend the argument for the general case and update the proof.
(iii) Showing is -Lipschitz instead of -Lipschitz, is possible. For , this follows from the above. We will attempt to tighten the proof in the next version of the draft.
- (i) and (ii) To verify the shortcut explanation, we have performed test-time evaluations on CIFAR-10, where we compare adversarial attacks with the correct label (Corr.: reported in the left 3 three columns) and predicted label (Pred.: reported in the right 3 three columns) and report the robust accuracy for various attack sparsities, of the SEA attack. We perform the analysis on three models (STD, SEA-AT, SEA-PGD AT) under signed rate encoding, while only SEA-AT originally had this anomaly. We also report the correct-to-incorrect (C -> I) and incorrect-to-correct (I -> C) rates, as given in the table below:
| Model | Corr. I→C | Corr. C→I | Corr. Acc. | Pred. I→C | Pred. C→I | Pred. Acc. | |
|---|---|---|---|---|---|---|---|
| Standard (clean 82.18) | 0 | 0.00 | 0.00 | 82.18 | 0.00 | 0.00 | 82.18 |
| 2 | 0.26 | 16.14 | 66.30 | 7.60 | 16.14 | 73.64 | |
| 4 | 0.10 | 26.09 | 56.20 | 8.88 | 26.09 | 64.97 | |
| 6 | 0.05 | 33.34 | 48.89 | 9.42 | 33.34 | 58.27 | |
| 8 | 0.05 | 38.57 | 43.66 | 9.57 | 38.57 | 53.18 | |
| 10 | 0.06 | 42.83 | 39.42 | 9.79 | 42.83 | 49.15 | |
| SEA-AT (clean 78.26) | 0 | 0.00 | 0.00 | 78.26 | 0.00 | 0.00 | 78.26 |
| 2 | 6.00 | 3.41 | 80.85 | 3.01 | 3.41 | 77.86 | |
| 4 | 7.50 | 3.29 | 82.46 | 2.83 | 3.29 | 77.80 | |
| 6 | 8.31 | 3.42 | 83.15 | 2.88 | 3.42 | 77.72 | |
| 8 | 8.56 | 3.82 | 83.01 | 2.94 | 3.82 | 77.38 | |
| 10 | 8.66 | 4.25 | 82.67 | 3.07 | 4.25 | 77.07 | |
| SEA-PGD-AT (clean 78.62) | 0 | 0.00 | 0.00 | 78.62 | 0.00 | 0.00 | 78.62 |
| 2 | 4.72 | 4.08 | 79.26 | 3.51 | 4.08 | 78.04 | |
| 4 | 5.77 | 4.74 | 79.65 | 3.51 | 4.74 | 77.39 | |
| 6 | 6.34 | 4.94 | 80.02 | 3.66 | 4.94 | 77.34 | |
| 8 | 6.44 | 5.60 | 79.46 | 3.89 | 5.60 | 76.91 | |
| 10 | 6.56 | 6.14 | 79.03 | 4.11 | 6.14 | 76.58 |
It can be observed that the anomaly can be avoided when predicted labels are used instead of correct labels while constructing the attack. However, using predicted label also amplifies the robust accuracy in other cases.
We thank the reviewer for the comments. Equation (14) indeed has a typo. The correct expression should be sign(g)−z∣sign(g)∣. We have verified that it does not affect the experiments, we will correct this in the next version of the draft.
Q1: In Table 1, the clean accuracy of a certain signed model in SVHN is 84.35%, but it increases to 93.05% after the SEA attack.
Reviewer J2Ut has the same question. Reproducing the reply for convenience.
As the number of non-zero spikes in Signed Rate Encoding is significantly lower than in Rate Encoding, an l0-norm-bounded perturbation leaves a stronger footprint on the encoded input under Signed Rate Encoding compared to Rate Encoding. It is possible that an adversarially trained model generates a spurious correlation between the adversarially perturbed image and the correct label, and when presented with the same attack, it performs better than without the perturbation. Since this is a shortcut, it actually disappears under the stronger SEA-PGD attack. The same phenomenon can be observed for CIFAR-10 (78.18-> 82.42) and CIFAR-100 (49.92 ->57.76).