Metric Learning for Robust High-Level Representations against Adversarial Attacks

Deep neural networks are powerful tools widely used for perception tasks. However, they are easily fooled by adversarial samples, which are created by adding imperceptible adversarial perturbations to clean samples. Input reconstruction is an effective method to resist adversarial samples without modifying the structure and parameters of the target model. However, input preprocessing can lead to a certain loss in the prediction accuracy of the protected model. In this paper, we introduce a novel input reconstruction approach incorporating metric learning within the high-level representation difference constraint. This aims to reduce the distance between samples of the same class while increasing the distance between different classes, thereby improving classification accuracy. To enhance the defense model’s transferability, it is no longer dependent on the target model, but instead, trained solely by the differences in high-level representations of a local classifier. Extensive experiments on three datasets demonstrate the superior performance of the proposed approach in defending against various attacks.