Pro-Trans: Progressive Tensor Ring with Attention Guided Local Smoothing Regularization

ICLR 2026 Conference Submission17456 Authors

Keywords: Tensor Networks, Adversarial Purification, Adversarial Attack, Tensor Ring

Abstract: The generalization of adversarial defense methods remains a critical open challenge, and optimization-based adversarial purification methods employing tensor network representations have recently shown strong potential. However, such tensor-based defense methods operate solely on the given input without relying on prior knowledge, which inevitably leads to overfitting to adversarial perturbations. Moreover, their iterative optimization procedures incur substantial computational overhead during inference. In this paper, we propose Pro-Trans, a novel tensor-based adversarial purification method that integrates progressive tensor ring with attention guided local smoothing regularization. Specifically, our progressive tensor ring avoids redundant upsampling operations, thereby reducing computational overhead and accelerating convergence. In addition, the proposed regularizer adaptively applies varying degrees of local smoothing regularization across different regions, thereby suppressing perturbations while mitigating semantic loss. Experimental results show that Pro-Trans consistently outperforms existing methods across diverse adversarial settings on CIFAR-10, CIFAR-100, and ImageNet, achieving state-of-the-art performance while maintaining low computational cost. The code will be available upon acceptance.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 17456