Towards Interpreting Vulnerability of Object Detection Models via Adversarial Distillation

Abstract

Recent works have shown that deep learning models are highly vulnerable to adversarial examples, limiting the application of deep learning in security-critical systems. This paper aims to interpret the vulnerability of deep learning models to adversarial examples. We propose adversarial distillation to illustrate that adversarial examples are generalizable data features. Deep learning models are vulnerable to adversarial examples because models do not learn this data distribution. More specifically, we obtain adversarial features by introducing a generation and extraction mechanism. The generation mechanism generates adversarial examples, which mislead the source model trained on the original clean samples. The extraction term removes the original features and selects valid and generalizable adversarial features. Valuable adversarial features guide the model to learn the data distribution of adversarial examples and realize the model’s generalization on the adversarial dataset. Extensive experimental evaluations have proved the excellent generalization performance of the adversarial distillation model. Compared with the normally trained model, the mAP has increased by 2.17% on their respective test sets, while the mAP on the opponent’s test set is very low. The experimental results further prove that adversarial examples are also generalizable data features, which obeys a different data distribution from the clean data. Understanding why deep learning models are not robust to adversarial samples is helpful to attain interpretable and robust deep learning models. Robust models are essential for users to trust models and interact with the models, which can promote the application of deep learning in security-sensitive systems.