You Can't Fool All the Models: Detect Adversarial Samples via Pruning Models

Many adversarial attack methods have investigated the security issue of deep learning models. Previous works on detecting adversarial samples show superior in accuracy but consume too much memory and computing resources. In this paper, we propose an adversarial sample detection method based on pruned models and evaluate four different pruning methods. We find that pruned neural network models are sensitive to adversarial samples, i.e., the pruned models tend to output labels different from the original model when given adversarial samples. Moreover, the pruned model has an extremely small model size and computational cost. Based on the detection result, we further propose a simple but effective defense approach to identify the true label of the adversarial sample. Experiments show that, on average, four different pruning methods outperform the SOTA multi-model based detection method (64.15% and 73.70%) by 28.65% and 18.73% on CIFAR10 and SVHN, respectively, with significantly fewer models used. The FLOPs of our structured pruned model are only 49.41% and 25.62% of the original model. Our defense approach achieves 68.60% and 72.03% average classification accuracy on CIFAR10 and SVHN, exceeding other advanced defense methods.