Test-time Detection and Repair of Adversarial Samples via Masked Autoencoder

Training-time defenses, known as adversarial training, incur high training costs and do not generalize to unseen attacks. Test-time defenses solve these issues but most existing test-time defenses require adapting the model weights, therefore they do not work on frozen models and complicate model memory management. The only test-time defense that does not adapt model weights aims to adapt the input with self-supervision tasks. However, we empirically found these self-supervision tasks are not sensitive enough to detect adversarial attacks accurately. In this paper, we propose DRAM, a novel defense method to detect and repair adversarial samples at test time via Masked autoencoder (MAE). We demonstrate how to use MAE losses to build a Kolmogorov-Smirnov test to detect adversarial samples. Moreover, we use the MAE losses to calculate input reversal vectors that repair adversarial samples resulting from previously unseen attacks. Results on large-scale ImageNet dataset show that, compared to all detection baselines evaluated, DRAM achieves the best detection rate (82% on average) on all eight adversarial attacks evaluated. For attack repair, DRAM improves the robust accuracy by 6% ~ 41% for standard ResNet50 and 3% ~ 8% for robust ResNet50 compared with the baselines that use contrastive learning and rotation prediction.

Test-time Defense against Adversarial Attacks: Detection and Reconstruction of Adversarial Examples via Masked Autoencoder

Existing defense methods against adversarial attacks can be categorized into training time and test time defenses. Training time defense, i.e., adversarial training, requires a significant amount of extra time for training and is often not able to be generalized to unseen attacks. On the other hand, test time defense by test time weight adaptation requires access to perform gradient descent on (part of) the model weights, which could be infeasible for models with frozen weights. To address these challenges, we propose DRAM, a novel defense method to Detect and Reconstruct multiple types of Adversarial attacks via Masked autoencoder (MAE). We demonstrate how to use MAE losses to build a KS-test to detect adversarial attacks. Moreover, the MAE losses can be used to repair adversarial samples from unseen attack types. In this sense, DRAM neither requires model weight updates in test time nor augments the training set with more adversarial samples. Evaluating DRAM on the large-scale ImageNet data, we achieve the best detection rate of 82% on average on eight types of adversarial attacks compared with other detection baselines. For reconstruction, DRAM improves the robust accuracy by 6% ~ 41% for Standard ResNet50 and 3% ~ 8% for Robust ResNet50 compared with other self-supervision tasks, such as rotation prediction and contrastive learning.