A Survey on Backdoor Attack and Defense in Deep Learning: A Paradigm Shift from Data-Driven to Behavior-Driven Triggers

The security of deep learning models is a critical concern, with backdoor attacks posing a significant threat. Existing research has focused on the data-driven paradigm, where triggers are embedded in input data. However, the widespread use of model post-processing techniques, such as quantization and fine-tuning, has created a novel and more stealthy attack vector. This paper systematically charts the evolution of backdoor attacks to address this emerging threat. This paper employs the method of literature review to categorize classic data-driven attacks and formally define the emerging behavior-driven paradigm, with its variants classified based on the triggering model behaviors. The analysis reveals a significant paradigm shift from triggers in "data content" to triggers in "model behavior." A key finding is that behavior-driven backdoors remain dormant in the original model, allowing them to bypass standard pre-deployment security screenings. This poses a critical threat to the AI supply chain. Consequently, the development of behavior-aware defense frameworks is not just a next step, but an urgent priority for future research.