acceptodds
Under review as a conference paper at ICLR 2027

FAST ADVERSARIAL TRAINING NEEDS DEBIASING AND CORRECTION TO ESCAPE THE TRAP

Abstract

Fast adversarial training (FAT) reduces the cost of robust training through single- step perturbation generation but remains vulnerable to catastrophic overfitting (CO), an abrupt collapse in adversarial robustness. By systematically dissecting CO from two complementary perspectives, we uncover two key failure mechanisms. First, we uncover an implicit class-specific backdoor-like effect: the outer minimization repeatedly associates biased perturbations with their ground-truth labels, causing persistent class-specific patterns to evolve into predictive triggers. Second, we un- cover gradient drift along the often-overlooked post-initialization update trajectory, where the initialization gradient becomes severely misaligned with subsequent local loss-ascent directions. To address these failures, we propose DC-FGSM, which combines class-adaptive spatial debiasing to disrupt the accumulation of class-specific trigger patterns with drift-aware gradient correction to compensate for directional deviations. Experiments across multiple datasets and perturbation budgets demonstrate that DC-FGSM achieves state-of-the-art robustness among single-step adversarial training methods. Furthermore, our gradient correction mod- ule consistently improves a broad range of existing FAT baselines, demonstrating its plug-and-play generality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial defense, adversarial robustness, Class-Adaptive Spatial Transformations, Gradient Correction

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

The paper studies catastrophic overfitting (CO) in fast adversarial training and proposes DC-FGSM, which uses class-adaptive perturbation priors and gradient correction. Experiments show robustness gains, but novelty, causal support, and numerical consistency remain insufficient.

Critical strengths
  1. Addresses a practical limitation of efficient adversarial training.
  2. Favorable cross-architecture results and improvements when added to existing FAT methods.
  3. Evaluation covers multiple datasets, backbones, attacks, and checkpoints.
Critical weaknesses
  1. Insufficient novelty of the mechanistic interpretation. The trigger-based explanation substantially overlaps with LIET (ICCV 2025), which already studies transferable class-label information after CO to explain the effectiveness of initialization. Gradient drift also needs clearer differentiation from GradAlign and prior local-nonlinearity analyses.
  2. Figures 1, 2, and 6, along with the entropy and drift analyses, largely demonstrate phenomena that accompany CO rather than establish its causes. Hence, the causal claims should be reduced to a mechanistic hypothesis.
  3. Incomplete ablations. For instance, the PGD-2 experiment changes both initialization and the attack update, so it cannot isolate the effect of drift correction. DC-FGSM also uses JSD regularization, masking/random replacement, and historical momentum, but Figures 7–8 do not fully separate these factors.
  4. Verifiable numerical inconsistencies and missing reproduction details. For CIFAR-100 at ϵ\epsilon = 8/255 (best checkpoint), DC-FGSM reports 27.27% AA accuracy in Table 1 but 27.09% in Table 5. In Table 8, the eight listed accuracies for FGSM-MEP average to about 32.07%, not the reported 29.5%; the reported Avg. standard deviation of 8.89 for FGSM-RS also conflicts with the mean of its eight component standard deviations (about 1.15). Reproduction also requires missing details such as the JSD coefficient, policy temperature/update rate, buffer capacity, policy update frequency, and attack step sizes/restarts.
Questions for authors

The paper addresses an important problem and shows promising robustness gains, but its novelty overlaps with prior work, its causal evidence is insufficient, and its ablations, baselines, and numerical consistency are inadequate

13 comments

We thank the reviewer for the constructive feedback. We clarify the relationship to prior work, present the additional controls, and detail the numerical corrections and implementation settings.

Weakness 1 — Novelty of the Mechanistic Interpretation and Method

(1) Relationship to LIET.

LIET demonstrates the transferability of label information in perturbations after CO and uses class-specific initialization derived from gray images, together with nonuniform label smoothing, to mitigate this effect. Our analysis builds on this observation by investigating the emergence and reinforcement of the class-dependent signal during training. LIET

Our contribution has three connected aspects. First, our trigger-based interpretation identifies persistent, spatially distributed class-dependent patterns (Figures 1 and 3) and investigates how they become reinforced during training. Second, we examine how changes in the gradient along the attack trajectory weaken inner maximization and how repeated outer updates can reinforce class-biased perturbations. Third, we use class-adaptive spatial transformations and trajectory gradient correction to address these two parts of the proposed feedback process. Figure 3 and the ablations in Figure 7 compare the adaptive transformations with fixed initialization and fixed spatial transformations.

Robustness under standard perturbation constraints. Our method restricts initialization noise to [−ϵ,ϵ][-\epsilon,\epsilon] and projects the final perturbation onto the standard L∞L_\infty ball. We train without early stopping and evaluate stability through the final checkpoint. As shown in Table III, debiased initialization substantially delays CO under these constraints, while Table IV shows that gradient correction further improves final-checkpoint robust accuracy. The full training objective also includes JSD consistency regularization, whose contribution is quantified separately in Tables VII and IX.

N-FGSM uses stronger noise together with an unclipped training perturbation construction. LIET's supplementary implementation description also specifies omission of perturbation clipping at larger budgets. These training choices motivate the constrained comparison below. N-FGSM, LIET supplementary material, Section B

(2) Temporal and intervention evidence.

Tables I and II use FGSM-AT branches from the same epoch-5 checkpoint, with interventions beginning at epoch 6(TNT). On 512 fixed examples, we measure

Gap=1N∑i=1N[L(xi+δiPGD,yi)−L(xi+δiinner,yi)].\mathrm{Gap}=\frac{1}{N}\sum_{i=1}^{N}\left[\mathcal{L}(x_i+\delta_i^{\mathrm{PGD}},y_i)-\mathcal{L}(x_i+\delta_i^{\mathrm{inner}},y_i)\right].

PGD denotes PGD-10, and inner denotes the training-time attack. A larger Gap indicates a weaker inner attack; a value near zero indicates comparable strength. We also report gradient drift, PGD-10 accuracy, and HTARt=HTRt/HTR4\mathrm{HTAR}_t=\mathrm{HTR}_t/\mathrm{HTR}_4 relative to a fixed reference. Epochs are zero-indexed, and “Final” is the last checkpoint.

Table I. Controlled Interventions on CIFAR-10 at ϵ=16/255\epsilon=16/255

MethodMetricEpoch 5Epoch 6(TNT)Epoch 7Epoch 10Epoch 20Epoch 22Epoch 40Epoch 80Final
Baseline (CO at Epoch 7)HTAR (×)1.751.930.948.558.195.234.492.292.24
Gradient Drift (%)23.9740.5976.3092.71116.21112.15110.15107.04105.53
Inner-Max Gap0.1110.1832.12317.84261.35756.61969.83685.54782.574
Robust Acc. (%)23.9621.794.010.040.000.000.020.110.09
Random-Noise Intervention (CO at Epoch 22)HTAR (×)1.750.850.841.111.185.9714.547.735.79
Gradient Drift (%)23.979.688.9311.0221.1098.79102.94104.78105.50
Inner-Max Gap0.1110.0980.1120.1650.32013.12749.75476.67973.313
Robust Acc. (%)23.9624.0524.2626.1828.240.000.000.010.01
Debiasing InterventionHTAR (×)1.750.990.921.151.271.201.381.721.74
Gradient Drift (%)23.9710.229.2811.6315.1816.3720.5929.1731.74
Inner-Max Gap0.1110.1560.1820.2880.4890.6661.1162.9363.610
Robust Acc. (%)23.9621.0921.5123.6625.6525.7126.6221.1819.89
Debiasing + CorrectionHTAR (×)1.751.000.951.111.241.241.281.601.62
Gradient Drift (%)23.979.608.9810.7713.9515.2318.1024.9727.22
Inner-Max Gap0.1110.1350.1640.2480.4020.5370.8422.1282.599
Robust Acc. (%)23.9621.6322.4525.1627.4227.2628.5324.1322.90

Table I shows that the baseline and random-noise branches collapse at epochs 7 and 22, respectively; random noise therefore only delays the same failure process. Debiasing retains 19.89% final PGD-10 accuracy, while gradient correction raises it to 22.90% and reduces both drift and the Inner-Max Gap. These results support both interventions from a shared pre-CO state.

Table II. Batch-Level Dynamics Around Catastrophic Overfitting on CIFAR-10 at ϵ=16/255\epsilon=16/255

Epoch / StageMetricB0–47B48–96B97–145B146–194B195–243B244–292B293–341B342–390
Epoch 6 (Pre-CO)HTAR (×)1.190.991.341.110.970.951.311.20
Gradient Drift (%)11.6512.3111.2814.4515.6012.7111.2717.16
Inner-Max Gap0.0890.1030.1260.0690.0680.0570.1080.132
Robust Acc. (%)30.4226.2128.8430.9328.5227.9028.1128.78
Epoch 7 (CO Transition)HTAR (×)1.290.961.891.711.697.254.987.82
Gradient Drift (%)16.3820.6527.2228.7338.8558.7679.4897.28
Inner-Max Gap0.1200.0970.1820.1660.2841.2942.0813.565
Robust Acc. (%)27.3628.7626.1825.6523.4921.015.692.39
Epoch 8 (Post-CO)HTAR (×)5.752.874.156.086.4910.1014.3517.75
Gradient Drift (%)99.93103.67103.42105.11105.88109.82110.67111.79
Inner-Max Gap6.4412.6836.00611.47112.66114.77223.41620.279
Robust Acc. (%)0.670.180.240.320.290.180.160.30

Table II resolves the transition at the batch level: drift accumulates throughout epoch 7, the Inner-Max Gap and HTAR rise sharply in B244–292, and robust accuracy collapses in the next window. This ordering places drift accumulation before robustness collapse and trigger amplification.

(3) Constrained comparison with LIET.

We compare LIET, D-FGSM, and DC-FGSM on CIFAR-10 at ϵ=16/255\epsilon=16/255, with all initialization noise restricted to [−ϵ,ϵ][-\epsilon,\epsilon] and all final perturbations projected onto the standard L∞L_\infty ball. D-FGSM uses debiasing initialization , and DC-FGSM additionally applies gradient correction. This experiment evaluates the methods under the specified training constraint.

In Table III, the mean CO onset shifts from epoch 30 for LIET to epoch 79 for D-FGSM across the three reported seeds, a delay of 49 epochs. D-FGSM still undergoes CO in these runs. Its role is therefore to extend the stable training period, while the final-checkpoint results in Table IV evaluate the additional benefit of gradient correction.

Table III. CO Dynamics of LIET and D-FGSM under Standard L∞L_\infty Projection and ϵ\epsilon-Bounded Initialization Noise on CIFAR-10 at ϵ=16/255\epsilon=16/255

MethodEpochFGSM Acc. (%)PGD Acc. (%)Gradient Drift
LIET
CO = 29.67 epochsSeed 0/1/2: 31/28/30———
1024.9325.837.96
2030.3730.2011.28
4073.670.97102.88
8083.802.23101.97
Final89.2717.03101.78
D-FGSM
CO = 78.67 epochsSeed 0/1/2: 78/85/73———
1022.8023.607.72
2027.6028.3010.62
4032.7032.3013.87
8058.302.1097.80
Final86.2013.20100.48

Table IV reports AA accuracy of 16.66%, 21.75%, and 23.46% at the best checkpoint for LIET, D-FGSM, and DC-FGSM, respectively. At the final checkpoint, AA accuracy is 0.00% for LIET, 0.05% for D-FGSM, and 19.41% for DC-FGSM. Final PGD-50 accuracy is 5.24%, 4.60%, and 29.09%, respectively. These results show that debiasing improves the best-checkpoint robustness in this setting, while gradient correction substantially improves the robustness retained at the end of training.

Because these comparisons use the same perturbation bound and projection, they establish that the reported gains can be obtained within the prescribed constraint. Figures 3 and 7 provide the more specific evidence concerning class-dependent spatial structure and transformation choice.

Table IV. Best- and Final-Checkpoint Performance under Standard L∞L_\infty Projection and ϵ\epsilon-Bounded Initialization Noise on CIFAR-10 at ϵ=16/255\epsilon=16/255

CheckpointMethodCleanPGD-50CWAACRPMAAvg.
BestLIET62.3124.4120.1416.6616.6617.2726.24
D-FGSM (Ours)71.9132.6226.7721.7521.9723.1033.02
DC-FGSM (Ours)72.5434.0027.4623.4623.4924.6034.26
FinalLIET44.685.243.150.0012.977.2712.22
D-FGSM (Ours)82.174.603.430.0520.3111.1120.28
DC-FGSM (Ours)77.5129.0924.6119.4119.7421.3331.95

Avg. is the arithmetic mean of Clean and the five attack-specific accuracy columns. Accuracy values are percentages.

Reviewer_TnYHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DC-FGSM, which augments fast adversarial training (FAT) with class-adaptive spatial debiasing and drift-aware gradient correction to mitigate catastrophic overfitting (CO). The authors analyze CO from two complementary perspectives: class-specific trigger-like patterns and gradient drift. They argue that repeated associations between perturbation patterns and ground-truth labels encourage the model to exploit these patterns as shortcuts for classification. Such patterns can increase the predicted probability of their associated class even when the perturbations are added to images from other classes. Gradient drift refers to the mismatch between the input gradient at initialization and local loss-ascent directions along the subsequent perturbation update, making the initial gradient less reliable for single-step adversarial example generation.

To address these CO traps, the authors propose two components:

  • Class-adaptive spatial debiasing to mitigate the trigger-like patterns: It breaks the repetitive perturbation by 1) making the representative perturbation pattern for each class, and 2) applying spatial transformations to these patterns. As the transformation may weaken the perturbations’ ability to confuse the model, it evaluates how adversarial they are with the pre-calculated gradients, and use the transformation that has high scores more frequent.
  • Drift-aware gradient correction to mitigate gradient drift: It checks how the gradient changes by 1) taking a small probing step from the debiased initialization along the sign of the initial gradient, and 2) calculating the gradient again at that point. As the initial gradient may become less reliable after moving the input, it uses the difference between the two gradients, scaled by a correction coefficient, to correct the initial gradient. It then generates the final adversarial sample with a single step using the sign of the corrected gradient. This final step starts from the original initialization, rather than continuing from the probing point.

The model is trained on the generated adversarial examples with an additional Jensen–Shannon divergence (JSD) regularizer that encourages prediction consistency between the debiased initialization and the final adversarial example. The authors evaluate DC-FGSM on CIFAR-10, CIFAR-100, and Tiny-ImageNet under multiple perturbation budgets and adversarial attacks, including PGD and AutoAttack. They report improved adversarial robustness over the compared FAT baselines, particularly in retaining robustness at the final checkpoint under larger perturbation budgets. Ablation studies examine the effects of adaptive spatial transformation selection and gradient correction. They also show that adding the gradient correction module to seven existing FAT methods generally improves robustness, suggesting that it can be used beyond DC-FGSM.

Critical strengths
  • Gradient correction can also benefit existing FAT methods: The authors apply the gradient correction module to seven existing FAT methods in Table 3 and show generally improved adversarial robustness on CIFAR-10 and CIFAR-100. For example, the AutoAttack accuracy of FGSM-RS on CIFAR-10 increases from 47.29% to 49.51%. This suggests that the correction module is useful not only within DC-FGSM, but also as an additional component for existing methods.

  • Better final-checkpoint robustness under larger perturbation budgets: The method shows improvements not only at the best checkpoint, but also at the end of training. In Table 1, DC-FGSM retains 19.41% AutoAttack accuracy on CIFAR-10 under ϵ=16/255\epsilon=16/255, compared with 15.80% for N-FGSM, while several other FAT baselines fall to nearly zero. This supports the method’s ability to mitigate robustness collapse during training, which directly addresses the CO problem.

Critical weaknesses
  • The causal explanation of CO needs stronger support: Figure 1 shows trigger-like patterns after CO, while Figure 2 shows gradient drift increasing alongside robustness collapse. These observations are consistent with the authors’ explanation, but they do not clearly distinguish whether these phenomena cause CO or emerge as consequences of it. Although the proposed components improve robustness, this improvement alone does not establish the claimed causal mechanisms. More controlled experiments separating changes in trigger persistence or gradient drift from changes in attack effectiveness would make the explanation more convincing.

  • The benefit of gradient correction is not fully separated from its additional computation: Gradient correction requires an extra forward–backward pass per batch when active. Therefore, the D-FGSM versus DC-FGSM comparison in Figure 8 changes both the attack-generation rule and the computational budget. A more controlled comparison would keep the debiased initialization, JSD regularization, and training schedule fixed, and compare the proposed correction with a standard two-step attack under the same gradient-evaluation budget. This would help clarify whether the correction rule itself provides a better use of the additional computation.

Questions for authors

Clarification of the causal interpretation: The analyses and ablations provide supporting evidence for the proposed explanation, such as reduced trigger responses in Figure 3 and reduced gradient drift with improved robustness in Figure 8. However, Figure 1 examines trigger-like behavior after CO, and Figure 2 shows gradient drift alongside robustness collapse. Could the authors clarify which results support these phenomena as causal contributors to CO, rather than only consequences or accompanying symptoms? In particular, how do the existing analyses and ablations distinguish the proposed mechanisms from other effects of changing the initialization or attack-generation procedure? This would help me assess the strength of the paper’s mechanistic contribution.

5 comments

We thank the reviewer for the constructive feedback. We clarify the proposed mechanism and explain how the temporal measurements, interventions, and ablations support it.

Weakness 1 and Question 1 — Evidence Supporting the Proposed CO Mechanism

(1) Relationship between gradient drift, inner-attack failure, and trigger-like behavior.

Our interpretation is that accumulated trajectory gradient drift weakens the training-time inner attack. The outer optimization then repeatedly fits class-biased perturbations, reinforcing persistent spatial patterns associated with particular classes. This feedback process provides a possible explanation for the trigger-like responses observed around and after catastrophic overfitting (CO).

Figure 2 examines the reliability of the inner-attack gradient, Figure 8 evaluates interventions that suppress or correct drift, and Figure 1 characterizes the resulting trigger-like behavior. Tables I and II provide additional intervention and temporal evidence. Together, these analyses support gradient drift as a contributor to CO and motivate examining how the inner and outer optimization interact.

(2) Intervention protocol and measurements.

All intervention branches start from the same FGSM-AT checkpoint at the end of the fifth training epoch and activate their respective interventions in the sixth epoch(INT). This protocol controls the model state and training history before intervention. We compare continued baseline training, random-noise injection, debiasing initialization , and debiasing initialization with gradient correction.

We evaluate inner-maximization effectiveness on a fixed set of N=512N=512 randomly selected examples using

=\frac{1}{N}\sum_{i=1}^{N} \left[ \mathcal L(x_i+\delta_i^{\mathrm{PGD}},y_i) -\mathcal L(x_i+\delta_i^{\mathrm{inner}},y_i) \right],$$ where $\delta_i^{\mathrm{PGD}}$ is generated by PGD-10 and $\delta_i^{\mathrm{inner}}$ by the training-time inner attack. A larger positive gap indicates that PGD-10 finds higher-loss perturbations than the inner attack. A gap near zero indicates similar mean loss under the two attacks. We also report gradient drift, PGD-10 robust accuracy, and the hard-trigger amplification ratio (HTAR), defined relative to a fixed reference hard-trigger response, $$\mathrm{HTAR}_{t}=\frac{\mathrm{HTR}_{t}}{\mathrm{HTR}_{4}}.$$ Epoch labels in the table follow zero-based training-epoch counts, and “Final” denotes the final checkpoint of the reported run. **Table I. Controlled Interventions on CIFAR-10 at $\epsilon=16/255$** | Method | Metric | Epoch 5 | Epoch 6(INT) | Epoch 7 | Epoch 10 | Epoch 20 | Epoch 22 | Epoch 40 | Epoch 80 | Final | | :--- | :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | Baseline **(CO at Epoch 7)** | HTAR (×) | 1.75 | 1.93 | 0.94 | **8.55** | 8.19 | 5.23 | 4.49 | 2.29 | 2.24 | | | Gradient Drift (%) | **23.97** | **40.59** | **76.30** | 92.71 | 116.21 | 112.15 | 110.15 | 107.04 | 105.53 | | | Inner-Max Gap | 0.111 | **0.183** | **2.123** | 17.842 | 61.357 | 56.619 | 69.836 | 85.547 | 82.574 | | | Robust Acc. (%) | 23.96 | **21.79** | **4.01** | 0.04 | 0.00 | 0.00 | 0.02 | 0.11 | 0.09 | | Random-Noise Intervention **(CO at Epoch 22)** | HTAR (×) | 1.75 | 0.85 | 0.84 | 1.11 | **1.18** | **5.97** | 14.54 | 7.73 | 5.79 | | | Gradient Drift (%) | 23.97 | 9.68 | 8.93 | 11.02 | **21.10** | **98.79** | 102.94 | 104.78 | 105.50 | | | Inner-Max Gap | 0.111 | 0.098 | 0.112 | 0.165 | **0.320** | **13.127** | 49.754 | 76.679 | 73.313 | | | Robust Acc. (%) | 23.96 | 24.05 | 24.26 | 26.18 | **28.24** | **0.00** | 0.00 | 0.01 | 0.01 | | Debiasing Intervention | HTAR (×) | 1.75 | 0.99 | 0.92 | 1.15 | 1.27 | 1.20 | 1.38 | 1.72 | **1.74** | | | Gradient Drift (%) | 23.97 | 10.22 | 9.28 | 11.63 | 15.18 | 16.37 | 20.59 | 29.17 | **31.74** | | | Inner-Max Gap | 0.111 | 0.156 | 0.182 | 0.288 | 0.489 | 0.666 | 1.116 | 2.936 | **3.610** | | | Robust Acc. (%) | 23.96 | 21.09 | 21.51 | 23.66 | 25.65 | 25.71 | 26.62 | 21.18 | **19.89** | | Debiasing + Correction | HTAR (×) | 1.75 | 1.00 | 0.95 | 1.11 | 1.24 | 1.24 | 1.28 | 1.60 | **1.62** | | | Gradient Drift (%) | 23.97 | 9.60 | 8.98 | 10.77 | 13.95 | 15.23 | 18.10 | 24.97 | **27.22** | | | Inner-Max Gap | 0.111 | 0.135 | 0.164 | 0.248 | 0.402 | 0.537 | 0.842 | 2.128 | **2.599** | | | Robust Acc. (%) | 23.96 | 21.63 | 22.45 | 25.16 | 27.42 | 27.26 | 28.53 | 24.13 | **22.90** | *Bold values highlight pre-collapse drift growth, CO transitions and associated trigger amplification, and paired final-checkpoint comparisons between debiasing alone and debiasing with gradient correction.*

(3) What the intervention and temporal results establish.

In Table I, baseline gradient drift rises from 23.97% at epoch 5 to 40.59% at epoch 6, before robust accuracy drops from 21.79% to 4.01% at epoch 7. The inner-maximization gap increases from 0.183 to 2.123 over the same transition. Random-noise injection postpones the collapse to epoch 22 in this run.

In the resumed debiasing branch, robust accuracy remains at 19.89% at the final checkpoint. Adding gradient correction raises it to 22.90% and reduces the final inner-maximization gap from 3.610 to 2.599. These results show that debiasing and correction improve the subsequent training trajectory from a shared pre-CO state.

The batch-level measurements in Table II provide complementary temporal evidence. During epoch 7, drift rises from 16.38% in B0–47 to 38.85% in B195–243, while robust accuracy decreases from 27.36% to 23.49%. The larger loss-gap increase and trigger amplification occur in B244–292, followed by the sharp accuracy drop to 5.69% in B293–341. This ordering supports drift accumulation before the abrupt robustness collapse, with trigger amplification developing during the transition.

Table II. Batch-Level Dynamics Around Catastrophic Overfitting on CIFAR-10 at ϵ=16/255\epsilon=16/255

Epoch / StageMetricB0–47B48–96B97–145B146–194B195–243B244–292B293–341B342–390
Epoch 6 (Pre-CO)HTAR (×)1.190.991.341.110.970.951.311.20
Gradient Drift (%)11.6512.3111.2814.4515.6012.7111.2717.16
Inner-Max Gap0.0890.1030.1260.0690.0680.0570.1080.132
Robust Acc. (%)30.4226.2128.8430.9328.5227.9028.1128.78
Epoch 7 (CO Transition)HTAR (×)1.290.961.891.711.697.254.987.82
Gradient Drift (%)16.3820.6527.2228.7338.8558.7679.4897.28
Inner-Max Gap0.1200.0970.1820.1660.2841.2942.0813.565
Robust Acc. (%)27.3628.7626.1825.6523.4921.015.692.39
Epoch 8 (Post-CO)HTAR (×)5.752.874.156.086.4910.1014.3517.75
Gradient Drift (%)99.93103.67103.42105.11105.88109.82110.67111.79
Inner-Max Gap6.4412.6836.00611.47112.66114.77223.41620.279
Robust Acc. (%)0.670.180.240.320.290.180.160.30

Bold values mark the transition from accumulated drift and inner-attack degradation to robustness collapse and the subsequent post-CO state.

Taken together, the temporal measurements and interventions support the proposed feedback mechanism. Because each intervention can affect several aspects of training, we interpret these findings as evidence for a causal contribution of trajectory drift, with the component controls below helping to assess the specificity of this interpretation.

Weakness 2 and Question 2 — Distinguishing the Proposed Mechanism from Other Effects

(1) How the analyses support different parts of the explanation.

Figures 1 and 3 characterize the class-dependent trigger response and its persistent spatial structure. Figure 2 examines changes in the attack gradient along the trajectory. Figure 8 evaluates debiasing initialization and gradient correction. The entropy analysis in Appendix F provides an additional description of representation changes during training. The intervention results in Tables I and II connect these observations to the evolution of inner-attack effectiveness and robust accuracy.

Our proposed roles for the two components are complementary. Class-adaptive spatial transformations disrupt recurring class-dependent perturbation patterns. Gradient correction uses the residual gp−g0g_p-g_0 to adjust the attack direction when the initial gradient changes along the trajectory.

(2) Comparisons controlling initialization and input-gradient evaluations.

Table III compares PGD-2, D-PGD-2, and DC-FGSM. D-PGD-2 uses the debiasing initialization within a two-step PGD update. D-PGD-2 and DC-FGSM share the initialization, auxiliary components, and training schedule, with two input-gradient evaluations per inner attack. The comparison tests the use of the second gradient for a corrected update from the initialization point against its use for a second PGD ascent step.

At ϵ=16/255\epsilon=16/255, final-checkpoint PGD-50 accuracy increases from 9.25% with PGD-2 to 11.53% with D-PGD-2 and 16.46% with DC-FGSM. AA accuracy increases from 5.99% to 7.76% and 10.29%, respectively. At ϵ=8/255\epsilon=8/255, DC-FGSM again improves all reported robust-accuracy measures over D-PGD-2, while the PGD-2–D-PGD-2 differences are smaller and mixed at the final checkpoint. DC-FGSM (E50) retains similar robust accuracy with a shorter training time.

Table III. Initialization and Gradient-Correction Comparisons on CIFAR-100

Perturbation Budget / CheckpointMethodCleanPGD-50CWAACRPMATime (s)
ϵ=8/255\epsilon=8/255 / BestPGD-251.6832.3427.5626.2926.2926.644434.66
D-PGD-254.6832.5128.0326.7026.6926.944545.63
DC-FGSM53.3832.9428.3627.1127.1527.434241.71
DC-FGSM (E50)53.5733.1828.5327.2227.2427.523712.78
ϵ=8/255\epsilon=8/255 / FinalPGD-252.6232.0527.2125.8725.8626.264434.66
D-PGD-255.7631.7827.2725.8925.8826.134545.63
DC-FGSM54.5132.8428.1126.7326.7527.124241.71
DC-FGSM (E50)54.7632.8428.1326.7926.7827.153712.78
ϵ=16/255\epsilon=16/255 / BestPGD-257.1511.5610.478.418.298.654523.00
D-PGD-253.8214.2012.3510.1810.1310.544410.00
DC-FGSM47.6818.1914.8312.2212.5113.144174.25
DC-FGSM (E50)47.3818.4115.0011.9512.4113.053721.61
ϵ=16/255\epsilon=16/255 / FinalPGD-257.949.257.865.995.916.254523.00
D-PGD-255.6511.539.597.767.658.064410.00
DC-FGSM50.0616.4613.3810.2911.3312.004174.25
DC-FGSM (E50)50.5816.6713.6110.1611.2711.943721.61

Accuracy values are percentages. Time is the reported end-to-end training time.

(3) Specificity of the trajectory residual.

Table IV compares the true trajectory residual with random, spatially shuffled, and average-gradient alternatives under the same input-gradient budget. The true residual achieves the highest PGD-10 accuracy and lowest reported gradient drift at both checkpoints. At the final checkpoint, its PGD-10 accuracy exceeds the average-gradient control by 4.26 percentage points. This comparison supports the usefulness of the information retained in the trajectory residual for the tested correction rule.

Table IV. Ablation of Gradient Signals with a Matched Input-Gradient Budget on CIFAR-100 at ϵ=16/255\epsilon=16/255

CheckpointGradient SignalClean Acc. (%)FGSM Acc. (%)PGD-10 Acc. (%)Gradient Drift ↓
BestRandom Residual58.9043.108.370.9585
Shuffled Residual58.7350.806.570.9664
Average Gradient56.9340.7311.770.7948
True Trajectory Residual52.4729.1017.370.6140
FinalRandom Residual54.1720.8712.670.7204
Shuffled Residual46.5317.5311.370.7475
Average Gradient48.2320.7315.770.5873
True Trajectory Residual49.0324.6020.030.4814
Reviewer_9JsxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper argues that fast adversarial training is prone to failure due to two issues: the model learns fixed perturbation patterns as classification shortcuts, and the initial gradient of a single-step attack cannot accurately reflect the subsequent attack direction. To address these issues, it proposes DC-FGSM, which disrupts class-specific perturbation patterns through spatial transformations and performs an additional gradient probe, combining the resulting gradient with the initial gradient to correct the attack direction.

Critical strengths
  1. The paper proposes targeted improvements to address failure phenomena in fast adversarial training and validates their effectiveness through reasonably extensive experiments.
  2. The paper is clearly written.
Critical weaknesses

DC-FGSM obtains a more reliable attack direction through one additional gradient computation, to obtain a more reliable attack direction. This can be understood as a trade-off between the computational efficiency of single-step attacks and the effectiveness of multi-step attacks. I noticed that the appendix also compares the performance of DC-FGSM and PGD-2, and the proposed DC-FGSM outperforms PGD-2. I have two main questions:

  1. Why can DC-FGSM outperform PGD-2? Both use two input-gradient computations, but they differ in initialization and how the gradients are used. I would like a deeper mechanistic explanation, or controlled experiments that hold initialization, training objectives, and computational budgets constant to further clarify the sources of the performance gains.
  2. Can DC-FGSM itself be extended into a more general formulation? The “probe–correct–update” procedure could be repeated multiple times, forming a general framework in which the number of iterations controls the computational cost. The number of iterations could be treated as a hyperparameter adjusted according to the training budget, with the current method corresponding to the single-update case. In principle, the number of repetitions should not be part of the method itself, but rather a hyperparameter determined by the training cost budget. I suggest that the authors explore the relationship between performance and training cost across different numbers of iterations, fully investigate the method’s performance potential, and examine how the performance gains saturate.
Questions for authors

Please refer to the questions raised in the weaknesses above.

3 comments

We thank the reviewer for the constructive feedback. We clarify how DC-FGSM uses each gradient evaluation and the specific role it plays, and provide additional comparisons of robustness and training cost under the probe–correct–update framework.

Weakness 1 — Why DC-FGSM Outperforms PGD-2 Adversarial Training

(1) How the two methods use the second gradient.

Both PGD-2 and DC-FGSM use two input-gradient evaluations to construct the training perturbation. PGD-2 refreshes the gradient after its first ascent step and then takes a second step from the updated point. DC-FGSM uses the gradient change at a probe point to correct the direction of a final update anchored at the initialization point.

Let g(δ)=∇δL(x+δ,y)g(\delta)=\nabla_\delta\mathcal L(x+\delta,y), g0=g(δ0)g_0=g(\delta_0), and ΠS\Pi_{\mathcal S} denote projection onto the perturbation set. The PGD-2 update is

\qquad \delta_2=\Pi_{\mathcal S}\left[\delta_1+\alpha_2\operatorname{sign}(g(\delta_1))\right].$$ For DC-FGSM, let $\delta_p$ be the probe perturbation and $g_p=g(\delta_p)$. The trajectory residual and corrected update are $$r_{\mathrm{traj}}=g_p-g_0, \qquad \delta_{\mathrm{DC}}=\Pi_{\mathcal S}\left[\delta_0+\alpha\operatorname{sign}\left(g_0+\lambda r_{\mathrm{traj}}\right)\right].$$ The corrected gradient is $\tilde g=g_0+\lambda(g_p-g_0)$. **DC-FGSM uses $g_p-g_0$ as an attack-trajectory-conditioned correction signal, combines the two gradient vectors before applying the sign operation, and generates the final perturbation from $\delta_0$.** In contrast, PGD-2 applies the sign operation to each gradient separately and performs two successive ascent steps with an intermediate projection. Thus, under the same two-gradient budget, the key distinction is whether the second gradient corrects the initial attack direction or performs another standard ascent step. **Table I examines the choice of gradient signal within the correction framework.** The true trajectory residual achieves the highest PGD-10 accuracy and lowest reported gradient drift at both checkpoints. At the final checkpoint, PGD-10 accuracy is 20.03%, compared with 15.77% for the average-gradient control, 12.67% for the random-residual control, and 11.37% for the shuffled-residual control. These results support the value of the trajectory residual in this experiment. Tables II and III provide the direct comparisons with PGD-2 and D-PGD-2. True Trajectory Residual uses $g_0+\lambda r$, where $r=g_p-g_0$. Random and shuffled controls preserve the per-example $L_2$ residual norm using Gaussian noise or spatial shuffling shared across channels, respectively. Average Gradient uses $(g_0+g_p)/2$ without normalization. All variants use two input gradients, apply sign after combination, and share the initialization procedure, step size, and projection. **Table I. Ablation of Gradient Signals with a Matched Input-Gradient Budget on CIFAR-100 at $\epsilon=16/255$** | Checkpoint | Gradient Signal | Clean Acc. (%) | FGSM Acc. (%) | PGD-10 Acc. (%) | Gradient Drift ↓ | | :--------- | :--------------------------- | -------------: | ------------: | --------------: | ---------------: | | **Best** | Random Residual | 58.90 | 43.10 | 8.37 | 0.9585 | | | Shuffled Residual | 58.73 | 50.80 | 6.57 | 0.9664 | | | Average Gradient | 56.93 | 40.73 | 11.77 | 0.7948 | | | **True Trajectory Residual** | 52.47 | 29.10 | **17.37** | **0.6140** | | **Final** | Random Residual | 54.17 | 20.87 | 12.67 | 0.7204 | | | Shuffled Residual | 46.53 | 17.53 | 11.37 | 0.7475 | | | Average Gradient | 48.23 | 20.73 | 15.77 | 0.5873 | | | **True Trajectory Residual** | 49.03 | 24.60 | **20.03** | **0.4814** |

(2) Contributions of debiasing initialization and gradient correction.

We introduce D-PGD-2, which uses our class-adaptive debiasing initialization with a two-step PGD update. The PGD-2–D-PGD-2 comparison examines the effect of the initialization strategy. For the D-PGD-2–DC-FGSM comparison, we keep initialization, auxiliary components, and the training schedule fixed. Both methods use two input-gradient evaluations per inner attack. We also evaluate DC-FGSM (E50), which activates correction during the later stage of training to reduce the total training cost.

As shown in Tables II and III, DC-FGSM outperforms D-PGD-2 across all five robustness metrics at both checkpoints and perturbation budgets, supporting the effectiveness of gradient correction under the same two-gradient budget. At ϵ=16/255\epsilon=16/255 (Table III), PGD-50 accuracy increases from 14.20% to 18.19% at the best checkpoint and from 11.53% to 16.46% at the final checkpoint; final AA accuracy improves from 7.76% to 10.29%.

Debiasing improves all five robustness metrics at the best checkpoint under both budgets, with improvements extending to the final checkpoint at ϵ=16/255\epsilon=16/255 (Tables II and III). DC-FGSM (E50) further demonstrates a favorable efficiency–robustness trade-off, maintaining comparable robustness while reducing training time by 12.5% and 10.8% at ϵ=8/255\epsilon=8/255 and 16/25516/255, respectively.

Table II. Initialization and Gradient-Correction Comparisons on CIFAR-100 at ϵ=8/255\epsilon=8/255

CheckpointMethodCleanPGD-50CWAACRPMATime (s)
BestPGD-251.6832.3427.5626.2926.2926.644434.66
D-PGD-254.6832.5128.0326.7026.6926.944545.63
DC-FGSM53.3832.9428.3627.1127.1527.434241.71
DC-FGSM (E50)53.5733.1828.5327.2227.2427.523712.78
FinalPGD-252.6232.0527.2125.8725.8626.264434.66
D-PGD-255.7631.7827.2725.8925.8826.134545.63
DC-FGSM54.5132.8428.1126.7326.7527.124241.71
DC-FGSM (E50)54.7632.8428.1326.7926.7827.153712.78

Accuracy values are percentages. Time is the reported end-to-end training time.

Table III. Initialization and Gradient-Correction Comparisons on CIFAR-100 at ϵ=16/255\epsilon=16/255

CheckpointMethodCleanPGD-50CWAACRPMATime (s)
BestPGD-257.1511.5610.478.418.298.654523.00
D-PGD-253.8214.2012.3510.1810.1310.544410.00
DC-FGSM47.6818.1914.8312.2212.5113.144174.25
DC-FGSM (E50)47.3818.4115.0011.9512.4113.053721.61
FinalPGD-257.949.257.865.995.916.254523.00
D-PGD-255.6511.539.597.767.658.064410.00
DC-FGSM50.0616.4613.3810.2911.3312.004174.25
DC-FGSM (E50)50.5816.6713.6110.1611.2711.943721.61

Accuracy values are percentages. Time is the reported end-to-end training time.

Weakness 2 — Performance–Cost Trade-off Across Probe–Correct–Update Iteration Counts

We evaluate repeated Probe–Correct–Update cycles with k∈{1,2,4,8}k\in\{1,2,4,8\}. The original DC-FGSM uses k=1k=1. Increasing kk adds inner-optimization work, and the measured training time grows approximately linearly over the tested range.

As shown in Table IV, increasing kk from 1 to 2 on CIFAR-10 improves final-checkpoint AA accuracy from 20.18% to 21.27% at 1.44×1.44\times the training cost, while k=4k=4 or 8 yields no further improvement. We therefore use k=1k=1 as the efficiency-oriented default, with larger kk providing optional robustness gains at higher computational cost.

For the FGSM-RS variant in Table V, increasing kk from 1 to 2 raises final-checkpoint AA accuracy from 0.00% to 18.33% and PGD-50 accuracy from 1.47% to 25.05%. Further increases to k=4k=4 and 88 yield smaller gains. These results show that additional correction cycles can substantially improve the final robustness of this variant, with diminishing returns as computation increases.

Table IV. Effect of the Number of Probe–Correct–Update Cycles in DC-FGSM on CIFAR-10 at ϵ=16/255\epsilon=16/255

Dataset / CheckpointkkCleanPGD-50CWAACRPMAAvg.Time (s)Overhead
CIFAR-10 / Best174.9532.5326.5322.3622.4823.6833.764,149.231.00×
274.1633.9126.8723.1823.4424.1234.285,962.441.44×
476.4531.6225.4121.4321.4722.2633.119,647.062.33×
873.4334.3426.8723.2623.3824.0434.2217,056.534.11×
CIFAR-10 / Final177.4229.6524.9320.1820.4221.7032.384,149.231.00×
277.0630.5525.0421.2721.3422.0632.895,962.441.44×
476.8630.9224.9521.0921.2021.8632.819,647.062.33×
876.7230.6624.8021.0421.1321.9532.7217,056.534.11×

Avg. is the arithmetic mean of Clean and the five attack-specific accuracy columns. Accuracy values are percentages.

Table V. Effect of the Number of Probe–Correct–Update Cycles in the FGSM-RS Variant on CIFAR-10 at ϵ=16/255\epsilon=16/255

Dataset / CheckpointkkCleanPGD-50CWAACRPMAAvg.Time (s)Overhead
CIFAR-10 / Best150.5324.6320.0517.4217.5217.7524.653,289.621.00×
272.3630.5926.7222.2822.4923.1832.945,162.441.57×
472.2231.2127.0222.7523.1123.6733.338,940.062.72×
871.7831.4727.1323.0523.3223.8633.4416,394.534.98×
CIFAR-10 / Final177.341.470.780.001.923.4014.153,289.621.00×
276.7925.0523.0418.3318.6219.3230.195,162.441.57×
476.5226.5223.9219.2319.6220.1831.008,940.062.72×
876.3526.6323.9819.4619.8020.3431.0916,394.534.98×

Avg. is the arithmetic mean of Clean and the five attack-specific accuracy columns. Accuracy values are percentages.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.