Adversarial Robustness via Multi-experts Framework for SAR Recognition with Class Imbalanced

With the rapid development of deep learning technology, significant progress has been made in the field of synthetic aperture radar (SAR) target recognition algorithms. However, deep neural networks are vulnerable to adversarial attacks in practical applications, inducing learning models to make wrong predictions. Existing works on adversarial robustness always assumed that the datasets are balanced. While in real-world applications, SAR datasets always suffer serious imbalanced distributions, which brings challenges to target recognition tasks and also affects the adversarial defense of models. So far few works have been reported on adversarial robustness for imbalanced SAR target recognition. Besides, single model is easily limited to a specific adversarial sample distribution. Motivated by this, a multi-expert collaborative diagnosis framework based on Contrastive Self-Supervised Aggregation (CS2AME) is proposed. The framework trains multiple personalized expert models precisely for dealing with specific SAR targets by formulating three expert guidance schemes, to better deal with different adversarial samples. In addition, a contrastive self-supervised aggregation strategy is designed to adaptively aggregate the professional expertise of each expert model. Extensive adversarial robustness recognition experiments on three publicly available imbalanced SAR datasets have demonstrated that the proposed CS2AME outperforms existing works in terms of standard performance and robust performance.