计算机科学
人工智能
噪音(视频)
特征(语言学)
机器学习
鉴定(生物学)
钥匙(锁)
匹配(统计)
计算机视觉
模式识别(心理学)
作者
Wenjing Ruan,Yongjin Xian,Shihui Xue,Jing Zhao
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-21
卷期号:: 14097-14101
标识
DOI:10.1109/icassp55912.2026.11462059
摘要
With the rapid progress of generative models and the growing diversity of face forgery techniques, developing more robust DeepFake detectors has become critical. Although CLIP-based detectors excel at learning general visual representations, their reliance on global class tokens limits their ability to capture subtle local artifacts crucial for forgery detection. This paper introduces the Synergistic Anomaly-Informed Learning (SAIL) strategy, which enhances CLIP by incorporating local artifact sensitivity. We design an auxiliary branch with a deep artifact-aware head that locates regions where reconstruction fails due to inconsistent artifacts and emphasizes them through a soft weighting mechanism. The resulting perception-guided signals steer the main CLIP network to jointly learn global facial attributes and local artifact cues. This synergistic learning significantly improves anomaly awareness, enabling the detector to capture fine-grained forgery evidence while maintaining strong generalization across diverse manipulation types.
科研通智能强力驱动
Strongly Powered by AbleSci AI