Adversarial sample generation technology combining benign salient areas and improved GAN in white-box and black-box attack scenarios

A end-to-end malware visual security assessment and defense framework is proposed to address the serious vulnerability faced by malware visual classification models in combating escape attacks. Firstly, in a white box scenario, an end-to-end attack method based on benign salient region guidance (BREAM) is proposed, which achieves high concealment escape while maintaining the executable functionality of the code. Secondly, in black box scenarios, a generative attack technique was designed that combines multi bit perturbation strategies with an improved generative adversarial network. This network integrates residual upsampling, mask normalization, and self attention mechanisms. To resist advanced threats, a defense framework combining diffusion purification and contrastive learning (DPCL Defense) was further developed, which utilizes manifold projection to remove adversarial noise and reshape feature extraction capabilities. The results show that the success rate of BREAM's attack on ResNet-50 in white box scenarios is as high as 95.57% ± 0.8%. The improved GAN achieved a success rate of 91.8% ± 0.85% in black box attacks against EfficientNet-B0. Faced with high-intensity attacks, the classification accuracy of DPCL-Defense remains as high as 94.6% ± 0.60%, and the success rates of white box and black box attacks are significantly suppressed to 2.85% ± 0.88% and 3.5% ± 0.67%, respectively. The research effectively balances defense robustness and classification accuracy, providing solid technical support for building reliable intelligent network defense systems.