DINO-Detect: A Simple and Effective Framework for Blur-Robust AI-Generated Image Detection

DINO-Detect uses teacher-student knowledge distillation to improve the robustness of AI-generated image detection under motion blur: a DINOv3 teacher model trained on clear images is frozen, and its image features and logit responses are distilled to a student model trained on blurred images, enabling the student to output consistent representations under motion degradation. The authors state that the method achieves SOTA under both motion-blurred and clear conditions.