Poisoning-Based Backdoor Attacks in Computer Vision

Recent studies demonstrated that the training process of deep neural networks (DNNs) is vulnerable to backdoor attacks if third-party resources (e.g., samples) are adopted. Specifically, adversaries intend embed hidden backdoors into DNNs, where can be activated by pre-defined trigger patterns and leading malicious model predictions. My dissertation focuses on poisoning-based in computer vision. Firstly, I study propose more stealthy effective against image classification tasks both physical digital spaces. Secondly, reveal threats visual object tracking, which representative critical video-related tasks. Thirdly, explore how exploit as watermark techniques for positive purposes. design a Python toolbox (i.e., BackdoorBox) implements advanced defenses under unified flexible framework, based provide comprehensive benchmark existing methods at end.