Pith. sign in

REVIEW 4 major objections 6 minor 36 references

TriGuard: Testing Model Safety with Attribution Entropy, Verification, and Drift

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Attribution drift exposes a failure mode that adversarial accuracy misses in neural networks.

desk verdict Composition of existing safety checks plus an unnormalized, unvalidated drift metric; the internal numbers don't line up, so the headline claims don't yet hold. read the letter →

arxiv 2506.14217 v1 pith:OPNZ5Y2J submitted 2025-06-17 cs.LG cs.AI

classification cs.LGcs.AI
keywords modelsafetyattributiondriftinterpretabilityformalverificationIntegratedGradientsentropyregularizationadversarialrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that model safety cannot be read off accuracy or adversarial robustness alone, because a model can pass formal verification and still change its reasoning pattern under small changes to how explanations are computed. To show this, it introduces TriGuard, a framework that combines formal robustness checks, the entropy of Integrated Gradients saliency maps, and a new Attribution Drift Score that measures how much saliency changes between two reference baselines. The paper reports that drift and adversarial error are effectively uncorrelated across five architectures and three datasets, meaning drift captures a distinct kind of fragility. It further claims that adding a gradient-entropy penalty to training lowers drift without hurting accuracy, pointing to a practical way to make explanations more stable.

What carries the argument

TriGuard's load-bearing piece is the Attribution Drift Score (ADS), the L2 distance between two Integrated Gradients attribution vectors computed from different baselines (zero versus blurred versions of the input), which the paper treats as a proxy for explanation stability. The other two axes are formal verification via PGD search plus interval bound propagation (with CROWN-IBP as a certified baseline), and attribution entropy, the Shannon entropy of the normalized absolute Integrated Gradients. The training-side mechanism is an entropy-regularized loss, $L_{\text{total}} = L_{\text{CE}} + \lambda \cdot L_{\text{entropy}}$, where $L_{\text{entropy}}$ penalizes the entropy of the normalized input gradient, encouraging sparse and stable saliency.

What would settle it

Compute ADS for a model with provably stable explanations—for example, a linear classifier on MNIST, where the attribution is exactly the weight vector and does not change under input perturbation—and show that the zero-versus-blurred baseline yields a nonzero drift that varies with the input; that would indicate the metric measures baseline sensitivity rather than reasoning instability. Alternatively, a controlled experiment where a model's decision rule is switched from one feature to another (e.g., a synthetic dataset) should show a spike in ADS at the switch; if it does not, the metric fails as a stability test.

Watch

Extended reading notes

Core claim

The central claim is that attribution-based signals are complementary to adversarial accuracy and that verified models can still exhibit unstable reasoning. The paper defines the Attribution Drift Score (ADS) as the L2 distance between Integrated Gradients maps computed from a zero baseline and a blurred baseline, and observes that this drift correlates only weakly with attribution entropy (r=-0.22) and not at all with adversarial error (r=0.00) across 18 model-dataset combinations. It also demonstrates that entropy-regularized training, which penalizes the entropy of the normalized input-gradient distribution, consistently reduces drift while maintaining comparable accuracy; for example, SimpleCNN on MNIST sees drift drop from 16.64 to 1.73 with λ=0.10. These results are offered as evidence that TriGuard uncovers 'silent brittleness'—models that look safe by standard metrics but shift their explanation logic.

Load-bearing premise

The Attribution Drift Score assumes that the L2 distance between attribution maps under two arbitrary baselines is a meaningful measure of explanation stability, and the paper never validates this metric against a known ground-truth change in a model's reasoning.

Editorial extensions

If this is right

  • Models that pass formal verification can still exhibit attribution drift, so certified robustness alone does not guarantee stable reasoning.
  • A low or zero adversarial error does not imply explanation stability, as shown by ResNet50 on CIFAR-10 and DenseNet121 on MNIST.
  • Adding a gradient-entropy penalty during training reduces Attribution Drift Scores without a meaningful accuracy drop.
  • Attribution entropy and drift provide orthogonal safety signals relative to adversarial metrics, since drift and adversarial error are uncorrelated.
  • TriGuard's three-axis evaluation can be applied across architectures (SimpleCNN, ResNet, MobileNet, DenseNet) and datasets to rank models by safety rather than accuracy alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same contrastive-drift idea could be applied to other attribution methods (SmoothGrad, LIME, attention maps) to build a model-agnostic stability diagnostic.
  • If ADS is validated against known reasoning changes—for example, a model that provably relies on a spurious feature—it could become a standard audit metric for regulators; the paper does not yet provide such validation.
  • A testable extension is to use ADS as a training regularizer itself, not just an evaluation metric, or to combine it with adversarial training to see whether drift and robustness improve jointly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. TriGuard proposes a multi-axis safety evaluation framework that combines formal robustness verification (PGD plus a simplified interval bound propagation check), attribution entropy over normalized Integrated Gradients, and a new Attribution Drift Score (ADS) defined as the L2 distance between Integrated Gradients maps computed under two different baselines. The paper claims that verified models can still exhibit unstable reasoning, that ADS and entropy provide safety insights complementary to adversarial accuracy, and that entropy-regularized training reduces ADS without sacrificing accuracy. Experiments are reported for five architectures on MNIST, FashionMNIST, and CIFAR-10, and the code is released.

Significance. If the central claims were sound, TriGuard would be a useful step toward routine evaluation of explanation stability alongside robustness, and the paper addresses an important gap by attempting to combine verification, attribution concentration, and attribution stability. The release of code is a practical strength. However, the ADS metric is unnormalized and never validated against existing stability benchmarks, the formal verification claims are not backed by a sound certificate, and the internal results are inconsistent. Because the reported correlations and the regularization benefit both pass through ADS, the empirical contribution is not currently substantiated.

major comments (4)
  1. [Section 2.4] The Attribution Drift Score is defined as ADS = ||a(1) − a(2)||_2, an unnormalized L2 distance between Integrated Gradients maps under zero and blurred baselines. This conflates attribution magnitude with stability: for an approximately linear classifier, IG_i ≈ (x_i − b_i) g_i, so ADS ≈ ||(b(2) − b(1)) ⊙ g||_2, which scales with baseline difference, gradient magnitude, and input dimension. A training intervention that shrinks gradient magnitudes will therefore reduce ADS even if the relative saliency pattern is unchanged. The paper does not normalize ADS, does not compare it with existing stability measures (e.g., Ghorbani et al. 2019; Alvarez-Melis & Jaakkola 2018), and reports only one-model baseline sensitivity in Appendix A. Since the claims of orthogonality with adversarial error (§3.1) and the regularization benefit (§3.2) are both measured with ADS, the central empirical findings may be scale artifacts rather than safety signals.
  2. [Section 3.2 / Tables 3 and 4] The unregularized SimpleCNN/MNIST configuration is reported with drift 2.93 in Table 3 but with drift 16.64 in Table 4 (λ=0.00) and in the text of §3.2 and §4. Table 1 lists drift 2.48 for the same regularized model, and the text pairs the 16.64-to-1.73 drop with Table 4. The same configuration therefore receives different ADS values, and the claimed reduction from 16.64 to 1.73 is not consistent with the Table 3 value. Without a fixed evaluation protocol (baseline pair, normalization, random seed), the ADS values are not reproducible, which undermines the central regularization result.
  3. [Section 2.5 / Section 3.2 / Table 4] The entropy penalty in §2.5 is defined as the negative entropy of normalized input gradients, which should favor sparse, low-entropy gradient distributions. Yet the paper reports that entropy-regularized training increases attribution entropy (e.g., SimpleCNN/MNIST from 4.47 to 5.13 or 5.29 in Table 4) and produces 'more dispersed' saliency maps. This contradicts the stated mechanism that the regularizer 'encourages sparser and more localized attribution maps' and the Abstract's claim that it 'reduces explanation drift without sacrificing performance.' The authors need to reconcile the mechanism with the observed entropy increase or correct the description and experiments.
  4. [Section 2.2 / Table 1] The paper repeatedly refers to models as 'formally verified' or 'passing formal verification' (Abstract, §1, §3), but the verification axis in §2.2 is a simplified check that the predicted class dominates across interval bounds, without specifying how bounds are propagated through nonlinear layers, and the PGD search is heuristic. No soundness proof or completeness claim is given, and Table 1 shows that the simplified check is far more permissive than CROWN-IBP (e.g., DenseNet121 on MNIST passes FormaVerif but fails CROWN-IBP). The central observation that 'verified models can still exhibit unstable reasoning' is therefore unsupported, because the check does not constitute a rigorous certificate.
minor comments (6)
  1. [Section 2.7] The text states that TriGuard is evaluated across '18 model-dataset combinations,' but Table 1 reports 15 rows (5 models × 3 datasets); if the count includes additional runs or configurations, please clarify.
  2. [Section 1] There are grammatical errors in the use of 'axes' as a singular noun, e.g., 'Each axes offers partial insight' and 'each axes'; these should be 'axis.'
  3. [Table 3] Table 3 omits the CROWN-IBP column that appears in Table 1, so the reader cannot compare the certified robustness of unregularized and regularized models.
  4. [Section 2.5] The symbol ϵ is used both for the perturbation bound in §2.2 and for the numerical offset in the entropy formula in §2.5; please use distinct symbols to avoid ambiguity.
  5. [Section 2.6 / Table 1] The SmoothGrad2 column in Table 1 is not clearly defined: it is unclear whether the reported number is an entropy value, a drift score, or a distance; the definition in §2.6 should be linked to the reported metric.
  6. [Appendix B.2] Figure 6 is referenced in the Deletion/Insertion faithfulness evaluation but does not appear in the manuscript; all referenced figures should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TriGuard's empirical claims do not reduce by construction to its metric definitions, though the Attribution Drift Score lacks external validation.

full rationale

The paper's central claims are empirical rather than derivational. ADS is defined in Section 2.4 as the L2 norm of the difference between two Integrated Gradients maps computed from zero and blurred baselines. The finding that entropy-regularized training reduces ADS (Section 3.2 and Table 4) is not forced by the definition of either ADS or the regularizer. The entropy regularizer (Section 2.5) penalizes the entropy of the normalized input-gradient distribution; it does not directly minimize the baseline-dependent L2 distance, nor does it reference the zero or blurred baselines. Whether this training change reduces ADS is an empirical observation, not an equation-level tautology. Likewise, the reported correlations among entropy, drift, and adversarial error (Section 3.1) are computed from measured quantities, not derived from the metric definitions. The paper also includes external faithfulness benchmarks via Deletion/Insertion AUC (Appendix B.2), providing an independent anchor for attribution quality. There are no load-bearing self-citations; all cited works are external. The main weaknesses are validity and consistency concerns rather than circularity: ADS is unnormalized and no external benchmark validates it as a stability measure, and there is a numerical inconsistency (Table 3 lists SimpleCNN/MNIST unregularized drift as 2.93, while Table 4 and the text report 16.64 for the same configuration). These issues undermine confidence in the conclusions, but they do not make the derivation circular. No step in the paper reduces to its own input by construction, so the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central results depend on a handful of hand-chosen hyperparameters (lambda, epsilon, unreported PGD/IG details) and on the unvalidated assumption that a simple L2 distance between two IG maps captures explanation stability. No new physical or mathematical entities are introduced.

free parameters (4)
  • Entropy regularization coefficient lambda = 0.00, 0.01, 0.05, 0.10
    Chosen by hand; the headline drift reduction result is tuned to this sweep rather than derived.
  • Perturbation bound epsilon = 0.1 (MNIST/FashionMNIST), 0.3 (CIFAR-10)
    Adversarial accuracy and verification results depend entirely on this ad hoc choice; the CIFAR-10 value is unusually large.
  • PGD step count T and step size alpha = unspecified
    Not reported; the claimed empirical verification status depends on the attack budget, so the pass/fail results are not reproducible.
  • IG integration steps m and numerical offset delta = unspecified
    Not reported; entropy and drift values depend on these implementation details.
assumptions (4)
  • domain assumption Integrated Gradients provides faithful attributions for the tested models.
    Borrowed from Sundararajan et al. (2017); also questioned by later sanity checks (Kim et al. 2022), so attribution entropy/drift inherit this assumption.
  • ad hoc to paper The simplified IBP-style check (class dominance across input bounds) is a sound formal robustness certificate.
    Section 2.2 describes only input clipping and 'predicted class dominates across all bounds' with no bound-propagation details; this is not a verified sound method.
  • domain assumption Zero and blurred baselines are semantically meaningful references for assessing explanation stability.
    Section 2.4 motivates these choices qualitatively; the ADS values and all drift correlations depend on this pair.
  • ad hoc to paper Minimizing the entropy penalty in L_total encourages sparser attribution maps.
    Section 2.5 asserts sparsification, but Table 4 shows entropy increasing with lambda, so the stated mechanism contradicts the reported data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TriGuard: Testing Model Safety with Attribution Entropy, Verification, and Drift." pith.science (2026). https://pith.science/paper/OPNZ5Y2J

@misc{pith2026250614217,
  author       = {Pith},
  title        = {Pith review of: TriGuard: Testing Model Safety with Attribution Entropy, Verification, and Drift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OPNZ5Y2J}},
  note         = {Machine review of arXiv:2506.14217}
}
read the original abstract

Deep neural networks often achieve high accuracy, but ensuring their reliability under adversarial and distributional shifts remains a pressing challenge. We propose TriGuard, a unified safety evaluation framework that combines (1) formal robustness verification, (2) attribution entropy to quantify saliency concentration, and (3) a novel Attribution Drift Score measuring explanation stability. TriGuard reveals critical mismatches between model accuracy and interpretability: verified models can still exhibit unstable reasoning, and attribution-based signals provide complementary safety insights beyond adversarial accuracy. Extensive experiments across three datasets and five architectures show how TriGuard uncovers subtle fragilities in neural reasoning. We further demonstrate that entropy-regularized training reduces explanation drift without sacrificing performance. TriGuard advances the frontier in robust, interpretable model evaluation.

Figures

Figures reproduced from arXiv: 2506.14217 by the authors.

Figure 1
Figure 1. presents attribution entropy across all models and datasets. We observe that CIFAR-10 models consistently yield higher entropy, indicating more diffuse attributions, likely due to input complexity. Meanwhile, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Attribution drift across all models. FashionMNIST models show the highest drift variance, while MNIST models exhibit more stable saliency patterns. • Entropy and drift show a weak negative correlation (r = −0.22), suggesting that more focused attributions may yield slightly more stable explanations. • Entropy and adversarial error exhibit a moderate neg￾ative correlation (r = −0.45), indicating that entropy may part… view at source ↗
Figure 3
Figure 3. Integrated Gradients (IG) saliency on clean vs. adversarial inputs. Visual comparison for digit “7” shows that adversarial inputs yield noisy, displaced saliency — highlighting the need for stability-aware attribution metrics. tropy and lower drift scores while maintaining comparable accuracy. For instance, SimpleCNN on MNIST sees drift reduced from 16.64 to 1.73 with a slight increase in entropy (4.47 to 5.29) [PI… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Correlation plots between attribution metrics and adversarial error. Left: entropy vs. drift shows a weak negative correlation (r = −0.22). Middle: entropy vs. adversarial error shows a moderate negative correlation (r = −0.45). Right: drift vs. adversarial error shows…
Figure 5
Figure 5. Figure 5: Contrastive attribution map using IG under adversarial perturbation. Red and blue regions visualize attribution displace￾ment, capturing explanation instability even when the prediction remains correct. λ reduces drift significantly while maintaining accuracy. For exam…
Figure 6
Figure 6. Figure 6: Faithfulness evaluation via Deletion/Insertion curves for SimpleCNN (with vs. without entropy regularization). Regularized models show sharper decline (Deletion) and stronger recovery (Insertion), indicating improved saliency faithfulness. 12 [PITH_FULL_IMAGE:figures/…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages

  1. [1]

    and Jaakkola, T

    Alvarez-Melis, D. and Jaakkola, T. S. On the robustness of interpretability methods. In Advances in Neural Information Processing Systems (NeurIPS), 2018. URL https://arxiv.org/abs/1806.08049

  2. [2]

    A Multi-Policy Framework for Deep Learning-Based Fake News Detection

    Bastani, O., Lee, H., and Raghunathan, A. Attribution under distribution shift: Quantifying the drift of explanation. In International Conference on Learning Representations (ICLR), 2023. URL https://arxiv.org/abs/2206.11866

  3. [3]

    Transformer interpretability beyond attention visualization

    Chefer, H., Gur, S., and Wolf, L. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 782--791, 2021

  4. [4]

    When are saliency maps trustworthy? In International Conference on Learning Representations (ICLR), 2023

    Chen, Z., Wang, J., Song, Y., Zhang, X., Liu, Y., and Zhou, Y. When are saliency maps trustworthy? In International Conference on Learning Representations (ICLR), 2023

  5. [5]

    and Hein, M

    Croce, F. and Hein, M. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020

  6. [6]

    Training verified learners with learned verifiers

    Dvijotham, K., Stanforth, R., Gowal, S., Uesato, J., and Kohli, P. Training verified learners with learned verifiers. In Advances in Neural Information Processing Systems (NeurIPS), 2018. URL https://arxiv.org/abs/1805.10265

  7. [7]

    Fong, R. C. and Vedaldi, A. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE international conference on computer vision, pp.\ 3429--3437, 2017

  8. [8]

    AI2 : Safety and robustness certification of neural networks with abstract interpretation

    Gehr, T., Mirman, M., Drachsler-Cohen, D., Tsankov, P., Chaudhuri, S., and Vechev, M. AI2 : Safety and robustness certification of neural networks with abstract interpretation. In IEEE Symposium on Security and Privacy, 2018

Show all 36 references
  1. [9]

    Ghorbani, A., Abid, A., and Zou, J. Y. Interpretation of neural networks is fragile. In AAAI, 2019

  2. [10]

    J., Shlens, J., and Szegedy, C

    Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015. URL https://arxiv.org/abs/1412.6572

  3. [11]

    Improving robustness without sacrificing accuracy via learned data augmentation

    Gowal, S., Qin, C., Uesato, J., Mann, T., and Kohli, P. Improving robustness without sacrificing accuracy via learned data augmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2021

  4. [12]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., and Song, D. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICLR, 2021

  5. [13]

    A benchmark for interpretability methods in deep neural networks

    Hooker, S., Erhan, D., Kindermans, P.-J., and Kim, B. A benchmark for interpretability methods in deep neural networks. In NeurIPS, 2020

  6. [14]

    Benchmark for evaluating saliency methods

    Hooker, S., Erhan, D., Kindermans, P.-J., and Kim, B. Benchmark for evaluating saliency methods. In Advances in Neural Information Processing Systems, 2021

  7. [15]

    The complete verification of neural networks: the eran approach

    Katz, G., Wang, W., Shiraishi, S., et al. The complete verification of neural networks: the eran approach. In CAV, 2021

  8. [16]

    Sanity simulations for saliency methods

    Kim, J., Ghorbani, A., and Zou, J. Sanity simulations for saliency methods. In International Conference on Learning Representations (ICLR), 2022. URL https://arxiv.org/abs/2105.06506v3

  9. [17]

    T., Dähne, S., and Erhan, D

    Kindermans, P.-J., Hooker, S., Adebayo, J., Alber, M., Schütt, K. T., Dähne, S., and Erhan, D. The reliability of saliency methods. In NeurIPS, 2019

  10. [18]

    Towards deep learning models resistant to adversarial attacks

    Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations (ICLR), 2018. URL https://arxiv.org/abs/1706.06083

  11. [19]

    Causal tracing for understanding unsafe outputs in large language models

    Meng, K., Andonian, A., Bau, D., and Belinkov, Y. Causal tracing for understanding unsafe outputs in large language models. arXiv preprint arXiv:2305.13928, 2023

  12. [20]

    Centered kernel alignment losses for saliency method faithfulness

    Ramaswamy, V., Moayeri, M., and Lakkaraju, H. Centered kernel alignment losses for saliency method faithfulness. In ICLR, 2022

  13. [21]

    Ross, A. S. and Doshi-Velez, F. Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In AAAI Conference on Artificial Intelligence, 2018. URL https://arxiv.org/abs/1711.09404

  14. [22]

    Evaluating the visualization of what a deep neural network has learned

    Samek, W., Montavon, G., Lapuschkin, S., Anders, C., and Müller, K.-R. Evaluating the visualization of what a deep neural network has learned. In IEEE Transactions on Neural Networks and Learning Systems, volume 28, pp.\ 2660--2673, 2017

  15. [23]

    Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models

    Samek, W., Montavon, G., Lapuschkin, S., Anders, C., and Müller, K.-R. Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models. Proceedings of the IEEE, 110 0 (5): 0 680--707, 2022

  16. [24]

    Deep inside convolutional networks: Visualising image classification models and saliency maps

    Simonyan, K., Vedaldi, A., and Zisserman, A. Deep inside convolutional networks: Visualising image classification models and saliency maps. In ICLR Workshop, 2014

  17. [25]

    An abstract domain for certifying neural networks

    Singh, G., Gehr, T., Mirman, M., Püschel, M., and Vechev, M. An abstract domain for certifying neural networks. In POPL, 2019

  18. [26]

    and Feizi, S

    Singla, S. and Feizi, S. Understanding adversarial robustness with visualizations of loss landscapes. In Advances in Neural Information Processing Systems, volume 32, 2019

  19. [27]

    Smoothgrad: removing noise by adding noise

    Smilkov, D., Thorat, N., Kim, B., Viégas, F., and Wattenberg, M. Smoothgrad: removing noise by adding noise. In arXiv preprint arXiv:1706.03825, 2017

  20. [28]

    Axiomatic attribution for deep networks

    Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In ICML, 2017

  21. [29]

    Z., Lin, C.-J., and Hsieh, C.-J

    Xu, K., Zhang, H., Lin, S., Zhao, S., Wang, X., Xu, M., Kolter, J. Z., Lin, C.-J., and Hsieh, C.-J. Fast and complete neural network verification by parallel bound propagation. In Advances in Neural Information Processing Systems, 2021 a

  22. [30]

    Xu, K., Zhang, H., Wang, S., Lin, C.-J., and Kolter, J. Z. Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers. In International Conference on Learning Representations (ICLR), 2021 b

  23. [31]

    Xu, M., Zhang, R., Lin, J., Wang, Y., and Wang, X. H. Reliable attribution of adversarial training. In NeurIPS, 2020

  24. [32]

    On the faithfulness and reliability of saliency explanations

    Yeh, C.-K., Kim, B., Arik, S., Li, C.-L., and Pfister, T. On the faithfulness and reliability of saliency explanations. Transactions on Machine Learning Research (TMLR), 2023. Accepted with minor revisions

  25. [33]

    Efficient neural network verification with Auto-LiRPA : Towards scalable certified defense

    Zhang, H., Weng, T.-W., Yi, J., Lin, C.-J., and Daniel, L. Efficient neural network verification with Auto-LiRPA : Towards scalable certified defense. In Advances in Neural Information Processing Systems (NeurIPS), 2020

  26. [34]

    Z., Wang, S., and Weng, T.-W

    Zhang, H., Xu, K., Lin, X., Kolter, J. Z., Wang, S., and Weng, T.-W. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations (ICLR), 2021. URL https://arxiv.org/abs/1906.06316

  27. [35]

    Towards certified robustness of real-world neural networks

    Zhang, H., Weng, T.-W., Xu, H., Lin, C.-J., and Hsieh, C.-J. Towards certified robustness of real-world neural networks. In AAAI, 2022

  28. [36]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.