Abstract:Adversarial training integrates adversarial examples during model training to enhance robustness. However, its application in fixed dataset settings differs from real-world dynamics, where data accumulates incrementally. In this study, we investigate Adversarially Robust Class Incremental Learning (ARCIL), a method that combines adversarial robustness with incremental learning. We observe that combining incremental learning with naive adversarial training easily leads to a loss of robustness. We discover that this is attributed to the disappearance of the flatness of the loss function, a characteristic of adversarial training. To address this issue, we propose the Flatness Preserving Distillation (FPD) loss that leverages the output difference between adversarial and clean examples. Additionally, we introduce the Logit Adjustment Distillation (LAD) loss, which adapts the model's knowledge to perform well on new tasks. Experimental results demonstrate the superiority of our method over approaches that apply adversarial training to existing incremental learning methods, which provides a strong baseline for incremental learning on adversarial robustness in the future. Our method achieves AutoAttack accuracy that is 5.99\%p, 5.27\%p, and 3.90\%p higher on average than the baseline on split CIFAR-10, CIFAR-100, and Tiny ImageNet, respectively. The code will be made available.
Abstract:Adversarial training significantly improves adversarial robustness, but superior performance is primarily attained with large models. This substantial performance gap for smaller models has spurred active research into adversarial distillation (AD) to mitigate the difference. Existing AD methods leverage the teacher's logits as a guide. In contrast to these approaches, we aim to transfer another piece of knowledge from the teacher, the input gradient. In this paper, we propose a distillation module termed Indirect Gradient Distillation Module (IGDM) that indirectly matches the student's input gradient with that of the teacher. We hypothesize that students can better acquire the teacher's knowledge by matching the input gradient. Leveraging the observation that adversarial training renders the model locally linear on the input space, we employ Taylor approximation to effectively align gradients without directly calculating them. Experimental results show that IGDM seamlessly integrates with existing AD methods, significantly enhancing the performance of all AD methods. Particularly, utilizing IGDM on the CIFAR-100 dataset improves the AutoAttack accuracy from 28.06% to 30.32% with the ResNet-18 model and from 26.18% to 29.52% with the MobileNetV2 model when integrated into the SOTA method without additional data augmentation. The code will be made available.