Adversarial Perturbation
Adversarial perturbation research focuses on developing and mitigating the vulnerability of machine learning models to maliciously crafted inputs designed to cause misclassification or other errors. Current research emphasizes improving the robustness of various model architectures, including deep convolutional neural networks, vision transformers, and graph neural networks, often employing techniques like adversarial training, vector quantization, and optimal transport methods. This field is crucial for ensuring the reliability and security of AI systems across diverse applications, from image classification and face recognition to robotics and natural language processing, by identifying and addressing vulnerabilities to attacks.
Papers
Perturbation Towards Easy Samples Improves Targeted Adversarial Transferability
Junqi Gao, Biqing Qi, Yao Li, Zhichang Guo, Dong Li, Yuming Xing, Dazhi Zhang
Exploring Adversarial Robustness of Deep State Space Models
Biqing Qi, Yang Luo, Junqi Gao, Pengfei Li, Kai Tian, Zhiyuan Ma, Bowen Zhou
Enhancing Adversarial Transferability via Information Bottleneck Constraints
Biqing Qi, Junqi Gao, Jianxing Liu, Ligang Wu, Bowen Zhou
Compositional Curvature Bounds for Deep Neural Networks
Taha Entesari, Sina Sharifi, Mahyar Fazlyab
The Price of Implicit Bias in Adversarially Robust Generalization
Nikolaos Tsilivis, Natalie Frank, Nathan Srebro, Julia Kempe
Probabilistic Perspectives on Error Minimization in Adversarial Reinforcement Learning
Roman Belaire, Arunesh Sinha, Pradeep Varakantham