ML Security PapersFakeRadar: Probing Forgery Outliers to Detect Unknown Deepfake Videos
Zhaolun Li 1, Jichang Li 2,3,4, Yinqi Cai 3, Junye Chen 3, Xiaonan Luo 1, Guanbin Li 3,2,4, Rushi Lan 1,5
1 Guilin University of Electronic Technology
4 Guangdong Key Laboratory of Big Data Analysis and Processing
5 Guangxi Key Laboratory of Image and Graphic Intelligent Processing
Published on arXiv
2512.14601
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
FakeRadar outperforms existing deepfake detectors on cross-domain benchmarks by explicitly modeling and covering unseen forgery regions via synthesized outlier samples during training.
FakeRadar
Novel technique introduced
In this paper, we propose FakeRadar, a novel deepfake video detection framework designed to address the challenges of cross-domain generalization in real-world scenarios. Existing detection methods typically rely on manipulation-specific cues, performing well on known forgery types but exhibiting severe limitations against emerging manipulation techniques. This poor generalization stems from their inability to adapt effectively to unseen forgery patterns. To overcome this, we leverage large-scale pretrained models (e.g. CLIP) to proactively probe the feature space, explicitly highlighting distributional gaps between real videos, known forgeries, and unseen manipulations. Specifically, FakeRadar introduces Forgery Outlier Probing, which employs dynamic subcluster modeling and cluster-conditional outlier generation to synthesize outlier samples near boundaries of estimated subclusters, simulating novel forgery artifacts beyond known manipulation types. Additionally, we design Outlier-Guided Tri-Training, which optimizes the detector to distinguish real, fake, and outlier samples using proposed outlier-driven contrastive learning and outlier-conditioned cross-entropy losses. Experiments show that FakeRadar outperforms existing methods across various benchmark datasets for deepfake video detection, particularly in cross-domain evaluations, by handling the variety of emerging manipulation techniques.
Key Contributions
- Forgery Outlier Probing: dynamic subcluster modeling with cluster-conditional outlier generation to synthesize samples simulating unseen forgery artifacts in feature space
- Outlier-Guided Tri-Training: a tri-class (Real/Fake/Outlier) optimization strategy using outlier-driven contrastive loss and outlier-conditioned cross-entropy loss
- Leverages CLIP's deep feature priors to proactively probe distributional gaps between real videos, known forgeries, and unseen manipulations for improved cross-domain generalization
🛡️ Threat Analysis
Proposes a novel AI-generated content detection framework for deepfake videos. FakeRadar's primary contribution is detecting AI-manipulated video content (output integrity), using outlier synthesis to generalize to unseen forgery types — a novel forensic/detection architecture, not merely a domain application of existing tools.