OrdinalNet - Dynamic and Robust Backdoor Attacks

DNN are vulnerable to backdoor attacks that can manipulate their outputs, leading to incorrect results. We point out the limitations of current backdoor attack methods, which rely on static and fixed triggers, making them easily detectable by existing backdoor defenses. In response, we propose a novel backdoor attack approach based on image features. This method leverages image to create an ordinal network that captures the precise image structure. Triggers are then generated using the information from this ordinal network. Additionally, we introduce a regularization method to enhance the robustness of our backdoor attack method against model diagnose-based defences. Our experimental results demonstrate the effectiveness of our approach. When compared to traditional backdoor attack methods such as BadNet and Blend, our method achieves attack success rate exceeding 99% on datasets like CIFAR-10, Tiny-ImageNet, and CelebA. Notably, the accuracy on clean samples only experiences a marginal decrease of less than 1%. Furthermore, our approach showcases its generalizability across different neural network architectures.