DSCAT: Dual Smoothing-constrained Adaptive Step Adversarial Training

Adversarial training has attracted considerable attention due to its effectiveness in safeguarding models against adversarial attacks in recent years. Among these methods, adversarial training with learnable attack strategies has stood out for its strong theoretical foundation and remarkable empirical performance. In this paper, we propose a novel adaptive attack strategy based on a single influencing factor for adversarial example generation: Dual Smoothing Constrained Adaptive Step-size Adversarial Training (DSCAT). DSCAT employs an entropy-constrained mechanism to smooth the adversarial example generation process, while simultaneously utilizing a perturbation intensity judgment and clipping layer to further regularize the generated adversarial examples. During training, the interplay between step-size adjustment and perturbation clipping effectively mitigates the “lagging” decision phenomenon observed in learnable attack strategies. This not only facilitates a smoother training process but also ensures that the model can make precise judgments when handling data near the decision boundary. Experimental evaluations on benchmark datasets demonstrate that DSCAT significantly outperforms other popular methods, substantially enhancing model robustness under a variety of attack scenarios.