Scholarlink

With the advancement of face forgery methods, deepfake detection has become a major concern. Recent studies demonstrate outstanding performance in intra-domain scenarios. However, they inevitably encounter generalization challenges when evaluated on unseen data. This paper introduces YOLODF, a novel YOLO-based deepfake detection model that enhances generalizability and mitigates overfitting through optimizations at both the data and feature levels. Specifically, Random Erasing Mask (REM) alleviates model’s undue dependence on inherent data biases via occluding and adding noise in sensitive facial regions. Multi-scale Cross-domain Feature Interaction (MCFI) utilizes the high-pass filter to explore and fusion multi-scale anomalies in local regions. Moreover, Domain Invariant Clues Learning (DICL) strengthens the learning of intrinsic representations by reconstructing the low-frequency distribution and optimizing crucial local features. We extensively evaluate and compare YOLODF for both in-dataset and cross-dataset tests on five public datasets. Experiment results confirm that our method can achieve a state-of-the-art generalization performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

YOLODF: YOLO-Based Spatial-Frequency Interaction Mining for General Deepfake Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xin Li,
  • Bingxin Xu,
  • Hongzhe Liu,
  • Weiguo Pan,
  • Cheng Xu

摘要

With the advancement of face forgery methods, deepfake detection has become a major concern. Recent studies demonstrate outstanding performance in intra-domain scenarios. However, they inevitably encounter generalization challenges when evaluated on unseen data. This paper introduces YOLODF, a novel YOLO-based deepfake detection model that enhances generalizability and mitigates overfitting through optimizations at both the data and feature levels. Specifically, Random Erasing Mask (REM) alleviates model’s undue dependence on inherent data biases via occluding and adding noise in sensitive facial regions. Multi-scale Cross-domain Feature Interaction (MCFI) utilizes the high-pass filter to explore and fusion multi-scale anomalies in local regions. Moreover, Domain Invariant Clues Learning (DICL) strengthens the learning of intrinsic representations by reconstructing the low-frequency distribution and optimizing crucial local features. We extensively evaluate and compare YOLODF for both in-dataset and cross-dataset tests on five public datasets. Experiment results confirm that our method can achieve a state-of-the-art generalization performance.