Diffusion Models Demand Contrastive Guidance for Adversarial Purification to Advance

• Autores: Bai, Mingyuan; Huang, Wei; Li, Tenghui; Wang, Andong; Gao, Junbin; Caiafa, César Federico; Zhao, Qibin
• Año de publicación: 2024
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: stable diffusion; adversarial attacks; purification; artificial intelligence
• Institución: Consejo Nacional de Investigaciones Científicas y Técnicas

In adversarial defense, adversarial purification can be viewed as a special generation task with the purpose to remove adversarial attacks and dif- fusion models excel in adversarial purification for their strong generative power. With different predetermined generation requirements, various types of guidance have been proposed, but few of them focuses on adversarial purification. In this work, we propose to guide diffusion mod- els for adversarial purification using contrastive guidance. We theoretically derive the proper noise level added in the forward process diffu- sion models for adversarial purification from a feature learning perspective. For the reverse pro- cess, it is implied that the role of contrastive loss guidance is to facilitate the evolution towards the signal direction. From the theoretical findings and implications, we design the forward process with the proper amount of Gaussian noise added and the reverse process with the gradient of contrastive loss as the guidance of diffusion models for adversarial purification. Empirically, exten- sive experiments on CIFAR-10, CIFAR-100, the German Traffic Sign Recognition Benchmark and ImageNet datasets with ResNet and WideResNet classifiers show that our method outperforms most of current adversarial training and adversarial purification methods by a large improvement.
Fil: Bai, Mingyuan. Riken. Center of Advanced Intelligence Project; Japón
Fil: Huang, Wei. Riken. Center of Advanced Intelligence Project; Japón
Fil: Li, Tenghui. Riken. Center of Advanced Intelligence Project; Japón
Fil: Wang, Andong. Riken. Center of Advanced Intelligence Project; Japón
Fil: Gao, Junbin. The University of Sydney; Australia
Fil: Caiafa, César Federico. Provincia de Buenos Aires. Gobernación. Comisión de Investigaciones Científicas. Instituto Argentino de Radioastronomía. Consejo Nacional de Investigaciones Científicas y Técnicas. Centro Científico Tecnológico Conicet - La Plata. Instituto Argentino de Radioastronomía; Argentina
Fil: Zhao, Qibin. Riken. Center of Advanced Intelligence Project; Japón
41st International Conference on Machine Learning
Viena
Austria
Carnegie Mellen University