acceptodds
Under review as a conference paper at ICLR 2027

Data-free Adversarial Training: Achieving Robustness without Data and Robust Teacher

Abstract

Deep neural networks are vulnerable to adversarial perturbations, while the most empirically reliable defenses (i.e., adversarial training and robustness distillation) typically require access to real training data and, in the latter case, an additional robust teacher. These extravagant assumptions hinder the deployment of existing defenses when sensitive data and robust teacher are unavailable, while only a naturally trained pretrained model can be accessed. To achieve adversarial robustness when only a naturally pre-trained model is available, we propose Data-Free Adversarial Training (DFAT), a robust post-training framework that requires neither real data nor a robust teacher. We theoretically show that a locally accurate student surrogate can reveal attack directions aligned with the natural teacher's local sensitivity, and that robust risk on real data can be controlled by optimizing teacher quality, robust anchoring error, and the discrepancy between synthetic and real distributions. Based on this, DFAT first performs data-free adversarial training to construct a natural student surrogate and learn a surrogate data distribution. It then uses the frozen natural teacher as a clean semantic oracle, and performs self-adversarial training on the student by maximizing and minimizing teacher-student output discrepancies over synthetic local neighborhoods. We further identify two additional challenges amplified in the data-free scenario, namely the fairness challenge and the generalization challenge, and propose to mitigate them from the perspectives of entropy reweighting and probability sparsity, respectively. To further mitigate the performance loss caused by surrogate distribution shift, we incorporate a test-time adaptation (TTA) module to further explore the self-adaptation potential of DFAT. Experiments and analyses demonstrate that DFAT achieves effective and reliable adversarial robustness in the practical setting with neither real data nor robust teacher.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Data-free, adversarial attack, adversarial defense, fairness, generalizability, probability sparsity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mB4oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies adversarial robustness when only a naturally pretrained model is available, without access to the original training data or a robust teacher. DFAT first learns a generator and a natural student through data-free distillation, then adversarially trains the student using the natural teacher’s predictions on synthetic data. Sample- and class-level generator reweighting target class-wise fairness, while negative-KL regularization against a frozen student copy is intended to promote probability sparsity and synthetic-to-real generalization. An optional DENT-based test-time adaptation stage further improves performance. The paper provides conditional theoretical analyses and evaluates the framework on CIFAR-10, CIFAR-100, and Tiny-ImageNet using several architectures and attacks.

Critical strengths

S1. The reported results demonstrate practically meaningful robustness under restrictive training access. On CIFAR-10 with ResNet-18, DFAT achieves 82.41% clean accuracy and 37.79% AutoAttack accuracy without test-time adaptation. Table 5 also reports an increase from 0.31% AutoAttack accuracy after natural pretraining to 29.89% with self-adversarial training and 37.79% with the complete method. These results suggest that the framework has practical potential beyond preserving natural accuracy, which is the strongest positive consideration in my assessment.

S2. The robust-risk decomposition provides a useful conditional analysis of the setting. Theorem 2 distinguishes teacher error, low-margin teacher predictions, robust anchoring on synthetic samples, and synthetic-to-real distribution discrepancy. This helps clarify the conditions under which a non-robust teacher could supervise a robust student, although the result does not establish that the proposed regularizers improve these quantities.

Critical weaknesses

W1. The contribution over the closest same-setting prior work is not established. DataFreeShield—Lee et al., DataFreeShield: Defending Adversarial Attacks without Training Data, ICML 2024—already studies learning robustness from a non-robust pretrained model without real training data. Its soft-guided objective includes clean-student/clean-teacher and adversarial-student/clean-teacher matching, which directly overlap with the first two terms of Eq. (11). Thus, the basic setting and central supervision strategy are established prior work.

Although DataFreeShield is cited, it is absent from the experimental comparisons. The paper instead primarily compares against methods with robust teachers, and its strongest reported comparisons additionally apply DENT only to DFAT. Meanwhile, nasty regularization is explicitly adapted from NT/NAT. Consequently, the current experiments do not establish the incremental value of the proposed combination over the closest alternative under matched training and inference conditions. A direct DataFreeShield comparison is central to assessing both novelty and significance.

W2. The proposed fairness mechanism has a mathematical interpretation problem, and its supporting ablation is internally inconsistent. Equation (8) minimizes ∑cwcpclog⁡pc\sum_c w_c p_c\log p_c, assigning smaller weights to harder classes with the stated intention of generating more samples from those classes. 11843_Data_free_Adversarial_Tr However, with fixed positive weights, minimizing this term over the probability simplex gives

$p_c^\star=\exp\left(-1-\frac{\nu}{w_c}\right), $

where normalization requires ν>0\nu>0 for K≥3K\geq3. Within this optimum, smaller weights therefore imply smaller class probabilities—the opposite of the claimed effect. This calculation isolates the regularizer rather than the complete generator objective, but it directly challenges the explanation for the proposed weighting rule.

The empirical evidence does not resolve this concern: Figures 4 and 7 display the same four sets of plotted values, yet Figure 4 labels them as baseline-versus-DFAT comparisons, whereas Figure 7 labels them as DERD accuracy and confidence measurements. These incompatible descriptions prevent Figure 4 from supporting the claimed fairness improvement. Establishing this contribution requires clarification of the implemented objective and correctly identified per-class results, potentially including revised experiments.

Questions for authors

Q1. Essential comparison. What is the substantive methodological advance over DataFreeShield beyond adding generator reweighting and an NT/NAT-derived regularizer? Please provide a comparison using the same natural teacher, student architecture, attack budget, and comparable synthesis/training resources, with test-time adaptation either absent from both methods or applied to both.

Q2. Implementation and evidence clarification. Is Eq. (8) implemented exactly as written, and are the difficulty-dependent weights detached during generator optimization? How does the implemented update increase the representation of difficult classes despite the isolated regularizer’s opposite preference? Please report the corresponding generated class frequencies and clarify which measurements Figures 4 and 7 actually represent.

Q3. Regularization mechanism. Please specify the discrepancy function, KL direction, and temperature used in Eqs. (11)–(12). If the frozen student copy satisfies Sa(x^)=F(x^)S^a(\hat x)=F(\hat x) and the divergences use matching conventions, the corresponding positive and negative KL terms reduce to a change in the distillation coefficients. What evidence distinguishes the proposed sparsity mechanism from this effect? Comparisons against coefficient-matched SAT and a simple entropy regularizer would materially affect my assessment of the claimed contribution.

Q4. Test-time defense evaluation. Appendix E.3 defines the adaptive objective as LCE+LDENT\mathcal L_{\mathrm{CE}}+\mathcal L_{\mathrm{DENT}}. How is this objective connected to the classification loss of the model after its test-time updates? Please clarify whether the attack differentiates through adaptation or uses a justified approximation, how it handles random augmentations and model resets, and whether attack convergence was checked. Adding the adaptation loss to the attack objective alone does not establish that the complete adaptive defense has been adequately attacked. This clarification is important because DENT contributes substantially to the reported gains.

Reviewer_Q4RdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the problem of improving adversarial robustness when neither the original training data nor a pretrained robust teacher is available, and only a naturally pretrained model can be accessed. To address this challenging setting, the authors propose Data-Free Adversarial Training (DFAT), a framework designed to enhance adversarial robustness without relying on real training data or robust teacher models. The proposed framework consists of three stages. First, a data-free knowledge distillation procedure is employed to generate synthetic training samples and initialize a student model. Second, the frozen natural teacher provides clean semantic supervision, while self-adversarial training is performed by maximizing and minimizing the output discrepancy between the teacher and student. Class-wise reweighting and probability sparsity regularization are further introduced to address robustness disparities across classes and improve generalization from synthetic data. Third, an optional test-time adaptation module based on DENT is incorporated to mitigate distribution shifts between synthetic training data and real test inputs. The paper also presents two theoretical results concerning the feasibility of using a distilled student as a surrogate for identifying adversarial directions and an upper bound on the student's robust generalization risk under the real data distribution. Experiments are conducted on CIFAR-10, CIFAR-100, and Tiny-ImageNet under various adversarial attack settings.

Critical strengths
  1. Improving adversarial robustness without access to real training data or a pretrained robust teacher is a practically relevant problem. This setting is particularly meaningful for applications involving restricted data access, privacy-sensitive environments, and limited computational resources.
  2. Beyond the basic data-free adversarial training objective, the paper considers several additional challenges, including class-wise robustness disparities, probability redundancy, and distribution shifts at test time. These components form a relatively comprehensive training and inference framework. 3.The experiments cover multiple standard image classification datasets and different model architectures. The authors report natural accuracy, adversarial robustness, worst-class accuracy, and class-wise performance disparities, together with evaluations under black-box and adaptive attacks. These results provide a useful empirical basis for examining the proposed method.
Critical weaknesses
  1. Figure 4 on page 9 and Figure 7 in the appendix on page 14 contain identical sets of four bar charts and statistical values, despite being presented as results from different experimental settings. Figure 7 illustrates the class-wise accuracy and prediction confidence of DERD, whereas Figure 4 relabels the same results as comparisons between the baseline and DFAT on clean and adversarial samples to support the claimed improvement in class-wise fairness. Consequently, Figure 4 cannot serve as valid evidence for the effectiveness of the proposed fairness reweighting strategy. 2."DataFreeShield: Defending Adversarial Attacks without Training Data" has already investigated adversarial robustness enhancement using only a non-robust pretrained model without access to real training data, which closely matches the central problem setting of this paper. However, the authors do not sufficiently clarify the essential methodological differences between DFAT and DataFreeShield, nor do they provide a direct experimental comparison.Therefore, the current evidence is insufficient to substantiate the paper's claims regarding methodological novelty and independent contributions. 3.The prediction smoothing term in Eq. (11) encourages the student's clean predictions to match the natural teacher's predictions on perturbed inputs. Since the natural teacher itself lacks adversarial robustness, the rationale behind this design requires further justification.
Questions for authors

1.Why do Figure 4 and Figure 7 contain identical sets of four subplots and statistical values despite describing different models and experimental conditions? Please clarify this discrepancy, provide the correct class-wise fairness ablation results. 2.What are the essential algorithmic contributions of DFAT compared with DataFreeShield, which also addresses adversarial robustness without real training data or a robust teacher? Could the authors provide a direct comparison under matched experimental settings? Additionally, to fairly isolate the contribution of test-time adaptation, could the authors evaluate competing methods such as FERD and DERD with the same DENT module and report the associated inference overhead? 3.Given the vulnerability of the natural teacher to adversarial perturbations, why is the prediction smoothing term in Eq. (11) expected to provide reliable supervision, and is its contribution independently validated?

Reviewer_S9t3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to realize Adversarial Training (AT) when neither real training data nor a robust teacher model is available. It first establishes two theoretical results: 1) when a distilled student locally approximates a natural teacher, the adversarial directions maximizing teacher-student disagreement align with the real local sensitivity directions of the teacher, and 2) the robust risk on the real data distribution can be bounded by teacher quality, robust anchoring error on synthetic data, and distribution discrepancy between real and synthetic data. Guided by these theories, the paper then proposes a Data-Free AT (DFAT) framework with only a naturally pre-trained model, consisting of vanilla self-distillation as pre-training, fair self-robustness learning, and an optional test-time adaptation.

Critical strengths
  • This paper explores a realistic AT setting that has neither real data nor a robust teacher.

  • The theoretical analysis looks promising.

  • The experimental setup regarding target models and attack scopes fits the standard practice in the AT field.

Critical weaknesses
  • There is a gap between the theoretical analysis and the actual experimental implementation. Theorem 1 requires a frozen copy S0S_0 that remains locally close to FF on perturbed inputs to guarantee valid attack directions. In Eq. 10, however, inner maximization is performed using the actively updated robust student. Also, Theorem 2 relies on a fixed generator producing i.i.d. samples, whereas Stage 2 continuously updates the generator online using student feedback.

  • The prediction smoothing term is questionable to me. Say, in Eq. 11, FF is a naturally trained model with basically no adversarial robustness, thus yielding bad predictions. In that case, forcing the clean student prediction to match that may not be reasonable.

  • The advantage of DFAT beyond the baselines looks to heavily rely on the test-time TTA. As it is basically an independent contribution, there would be a question: How about directly applying TTA on the baselines? If it also works to a large extent, that may imply the first two main stages of DFAT are not that effective (as by themselves DFAT is still behind the baselines in Tab. 1 and 2). On the other hand, if TTA is finally proven not useful for the baselines, its additional test-time cost compared with the baselines should then be considered.

  • The soundness and presentation of this paper are not good. I notice multiple inconsistencies across the method, equations, and settings. Representatively, Sec.1 and Fig. 2 state that adaptive weighted entropy and weighted one-hot losses are applied in Stage 1, but Sec. 2.2.1 formulates Stage 1 with unweighted losses, while introducing the weighted ones only in Stage 2. Also, Fig. 4 and 7 seem to misuse the same figure.

Questions for authors

See weaknesses pls.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.