Learning Better Certified Models from Empirically-Robust Teachers
Authors: Alessandro De Palma
Organizations: Department of Statistics, London School of Economics and Political Science, London, United Kingdom · Work parly carried out at Inria, École Normale Supérieure, PSL University, CNRS
Adversarial training attains strong empirical robustness to specific adversarial attacks by training on concrete adversarial perturbations, but it produces neural networks that are not amenable to strong robustness certificates through neural network verification. On the other hand, earlier certified training schemes directly train on bounds from network relaxations to obtain models that are certifiably robust, but display sub-par standard performance. Recent work has shown that state-of-the-art trade-offs between certified robustness and standard performance can be obtained through a family of losses combining adversarial outputs and neural network bounds. Nevertheless, differently from empirical robustness, verifiability still comes at a significant cost in standard performance. In this work, we propose to leverage empirically-robust teachers to improve the performance of certifiably-robust models through knowledge distillation. Using versatile feature-space distillation objectives, we show that distillation from adversarially-trained teachers can significantly improve the performance of state-of-the-art certified training algorithms for ReLU networks on robust computer vision benchmarks.
Figures & tables
Figure 1: Standard, empirical adversarial and certified accuracies (BaB and CROWN/IBP) under ℓ∞ perturbations of networks trained using CC-Dist under varying distillation coefficient β . The legend is reported once for all subfigures in plot 1(b) . Metrics are reported on the CIFAR-10 test set. The β value employed throughout the paper (see \lx@sectionsign 3.2 ) is marked by a dashed vertical line.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Table 10: Effect of distillation beyond CNN-7 students on CIFAR-10 with ϵ=\nicefrac2255 for CC-Dist.
Certified training aims to produce models whose predictions can be formally verified against adversarial perturbations, typically by optimising upper bounds on the worst-case loss over an allowed perturbation set. For neural networks, certified training methods based purely on tight relaxation bounds produce networks that are amenable to certification, but sacrifice standard accuracy. Conversely, adversarial training often yields stronger empirical robustness and standard accuracy, but the resulting models are generally difficult to certify with neural network verifiers. Recently, the literature has shown that better standard-certified accuracy trade-offs can be achieved by combining adversarial training objectives with loose over-approximations based on Interval Bound Propagation (IBP), effectively interpolating between lower and upper bounds of the worst-case loss. Building on this, we introduce AD-CERT, a certified training objective that combines adversarial distillation with an IBP upper bound. We show that distilling adversarial information over the logit space from an empirically robust teacher provides an effective lower bound surrogate for certified training, with AD-CERT achieving state-of-the-art certified performance on several robustness benchmarks for ReLU networks. Furthermore, in a unified setup, distilling adversarial information at the logit-level is shown to improve certified accuracy over a robust feature-space distillation objective by up to 5.40 percentage points. Our code has been made available at https://github.com/matteo-melis/ad-cert.
Matteo Melis, Jesus Martinez Del Rincon, Vishal Sharma
Adversarial distillation transfers robustness from high-capacity teachers to compact students. Existing adversarial distillation methods mainly use teacher predictions on clean or adversarial examples to supervise student learning. However, teacher-favorable supervision within the perturbation neighborhood remains underexplored in adversarial distillation. We therefore propose Collaboratively Guided Adversarial Robust Distillation (CGARD), which jointly optimizes distinct student-adversarial and teacher-collaborative examples within the same perturbation neighborhood. The teacher-collaborative example is constrained to incur no greater cross-entropy loss under the teacher than the clean input. CGARD combines collaborative teacher guidance with adversarial teacher supervision to improve robust knowledge transfer. Experiments on CIFAR-10 and CIFAR-100, including white-box evaluation and additional black-box transfer evaluation, demonstrate consistent robustness improvements over strong adversarial distillation baselines.
Zhi Li, Haowei Liu, Hongchen Yang +4
Engineering Research Center of Cyberspace and School of Software and AI, Yunnan University, Kunming, China · School of Information Science and Technology, Yunnan Normal University, Kunming, China
Adversarial Distillation aims to enhance student robustness by guiding the student with a robust teacher's soft labels within the min-max adversarial training framework, yet its success is notoriously inconsistent: a more robust teacher often fails to improve, or even harms, the student's robust generalization. In this paper, we identify a key mechanism of this teacher dependency: the misalignment between the teacher's supervisory confidence and the student's representational limitations on a consistent subset of training data -- the Robustly Unlearnable Set. We present a theoretical framework analyzing the feature learning dynamics of a two-layer neural network, demonstrating that this mismatch creates a dichotomy in distillation outcomes. We prove that when a teacher provides confident supervision on unlearnable samples, it compels the student to memorize spurious noise patterns that eventually overpower the learned robust signal, thereby driving robust overfitting. Conversely, a teacher that exhibits high uncertainty on these samples effectively suppresses noise memorization, allowing the student to rely solely on the learnable signal for robust generalization. We empirically validate our theory across both synthetic simulations and real-image classification datasets, confirming that robust overfitting is driven by the teacher's interaction with unlearnable samples. Finally, we demonstrate that a teacher's predictive entropy on unlearnable samples serves as a strong indicator of student robustness, validating our theoretical framework and offering a principled guideline for robust teacher selection.
Hongsin Lee, Hye Won Chung
School of Electrical Engineering, KAIST, Daejeon, Korea.