CCBA: Code Poisoning-Based Clean-Label Covert Backdoor Attack Against DNNs

Deep neural networks have been shown to be vulnerable to backdoor attacks, and currently, almost all attacks involve inserting backdoors into models through data poisoning, which requires the attacker to have access to higher-level model training and can be easily exposed. However, vulnerabilities in code management for deep learning training make the code itself an extremely susceptible target for attacks. based on this, we propose a novel form of backdoor attack called Code Poisoning-based Clean-Label Covert Backdoor Attack (CCBA), which dynamically modifies the training data by manipulating only a small fraction of the code to inject a backdoor. This attack imposes a negligible burden on the training process, while still achieving strong performance and maintaining stealth. We not only validate the feasibility and effectiveness of CCBA in deep neural networks but also extend it successfully to graph neural networks and natural language processing, demonstrating promising results.