论文学习笔记
DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection
一、研究动机
- Although large-scale multimodal models like CLIP offer strong transferable representations for detecting synthetic content, fine-tuning them often induces catastrophic forgetting, which degrades pre-trained priors and limits cross-domain generalization.
- The adapted model may overfit to dataset-specific patterns rather than learning artifact-level features that generalize across different generative mechanisms.
- Conventional feature distillation methods, which enforce global alignment between
teacher and student representations, may inadvertently retain these task-irrelevant factors.
二、研究目标
- Preserves transferable pre-trained priors while suppressing task-irrelevant components.
三、技术路线
