Symmetry-Breaking Dimensional Expansion: Improving Accuracy and Reducing Transferability under Transfer-Based Black-Box Attacks

This resource package presents an implementation of an adversarial defense mechanism based on symmetry-breaking dimension expansion, designed to simultaneously enhance clean accuracy and model robustness against transfer-based black-box adversarial attacks. The implementation constructs an ExpandedResNet architecture benchmarked on CIFAR-10 and ImageNet-100 datasets, incorporating spaced pixel expansion during the training phase, with comprehensive training and evaluation conducted under unified configuration settings.The package includes:CIFAR-10 data loading pipeline with standardized augmentation and normalization procedures;The expand_pixels dimension expansion operator and corresponding modified model architecture;Transfer attack evaluation scripts for measuring adversarial accuracy of victim models against multiple attack methodologies, including PGD, AutoAttack, TI-FGSM, MI-FGSM, BIM, and DeepFool;Pre-trained model weights and documented experimental results.This resource enables reproduction of experiments presented in our paper, including comparisons between standard natural training approaches, evaluations against state-of-the-art adversarial training methods, and ablation studies examining the impact of dimension expansion parameters. The implementation environment is built upon PyTorch and torchvision frameworks, with adversarial sample generation facilitated by the torchattacks library.