A Two-Stage Adversarial Training Method Based on Stability Contrastive Learning to Enhance Adversarial Robustness

Neural network models are highly susceptible to adversarial sample attacks, causing significant differences in model predictions with even minor perturbations to the samples. Adversarial training is a kind of effective technique for resisting sample adversarial attacks. Traditional adversarial training methods are all single-stage training, and in the middle and later stages of training, there is a serious issue of robust overfitting. The accuracy of the adversarial training model does not increase and may even experience severe degradation. For this issue, we propose a multi-stage stability contrastive adversarial training method based on Ulam stability, which performs adversarial training in an optimized space with stability constraints to improve the robustness accuracy and training stability of the model. In the first stage, a stability adversarial training strategy is adopted to enable the model to rapidly improve robust accuracy before overfitting. In the second stage, a stability contrastive learning strategy is employed, focusing on suppressing overfitting of the model and further enhancing robust accuracy. This two-stage adversarial training method can not only improve the robustness accuracy of the model, but also effectively suppress overfitting phenomena. This method has plug and play pendant coupling capability, which can be combined with existing multi-class models to further improve the robustness accuracy of the benchmark model. In addition, this method has the characteristic of stable convergence, which can reduce the dependence on early stopping conditions and make the model training more flexible. Comparative experiments on multiple datasets have also validated the effectiveness of the proposed method.