SDProtoL: Enhancing Rehearsal-Free Lifelong Face Forgery Detection via Prototype-Guided Prompt Tuning

The performance of deep neural networks in face forgery detection (FFD) is impressive, but they are obtained with static models incapable of adapting their behavior over time. However, in a dynamic world, FFD systems deployed on Internet of Things (IoT) devices operate on vast streams of face data with ever-changing distributions, leading to catastrophic forgetting and significant declines in detection performance. To address this issue, we propose a novel rehearsal-free domain lifelong learning (RF-DLL) framework for FFD, termed SDProtoL. This framework adopts the prompt-based incremental learning paradigm and steers prototypes to further mitigate catastrophic forgetting and improve generalization to unseen data. Specifically, to alleviate the forgetting of previous domains without using previous data, we employ a Gaussian mixture model (GMM) to derive GMM-based hierarchical static prototypes (GHSPs) with nonconvex and anisotropic characteristics, in order to fit the complex distribution of face forgery data. Furthermore, for better generalization, we propose the dynamic prototype-oriented asymmetric contrastive regularization (DPACR), which improves generalization ability for unseen data by accumulating and transferring previous knowledge. Simulating practical RF-DLL scenarios, we establish a challenging lifelong FFD (LFFD) benchmark and construct three protocols referencing real-world scenarios. Extensive experimental results demonstrate that our proposed method significantly alleviates catastrophic forgetting while exhibiting superior generalization performance in unseen domains.