The Alpha Blending Hypothesis: Compositing Shortcut in Deepfake Detection
Summary
This paper introduces the Alpha Blending Hypothesis, suggesting deepfake detectors primarily identify compositing artifacts rather than semantic anomalies. It proposes a method called BlenD that achieves superior cross-dataset generalization using real-only image augmentation with self-blended images.
View Cached Full Text
Cached at: 05/12/26, 07:27 AM
Paper page - The Alpha Blending Hypothesis: Compositing Shortcut in Deepfake Detection
Source: https://huggingface.co/papers/2605.10334
Abstract
Deepfake detectors primarily function as alpha blending searchers that identify compositing artifacts rather than semantic anomalies, with a proposed method achieving superior cross-dataset generalization through real-only image augmentation.
Recentdeepfake detectionmethods demonstrate improvedcross-dataset generalization, yet the underlying mechanisms remain underexplored. We introduce theAlpha Blending Hypothesis, positing that state-of-the-artframe-based detectorsprimarily function as alpha blending searchers; rather than learning semantic anomalies or specific generative neural fingerprints, they localize low-level compositing artifacts introduced during the integration of manipulated faces into target frames. We experimentally validate the hypothesis, demonstrating that deepfake detectors exhibit high sensitivity to the so-calledself-blended images(SBI) and non-generative manipulations. We propose the method BlenD that leverages a large-scale, diverse dataset of real-only facial images augmented with SBI. This approach achieves the best averagecross-dataset generalizationon 15compositional deepfake datasetsreleased between 2019 and 2025 without utilizing explicitly generated deepfakes during training. Furthermore, we show that predictions from explicit blending searchers and models resilient to blending shortcuts are highly complementary, yielding a state-of-the-artAUROCof 94.0% in anensemble configuration. The code with experiments and the trained model will be publicly released.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2605\.10334
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.10334 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.10334 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.10334 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
EMO-BOOST: Emotion-Augmented Audio-Visual Features for Improved Generalization in Deepfake Detection
This paper proposes Emo-Boost, a multimodal deepfake detection framework that leverages emotion cues (audio-visual emotion recognition) as high-level semantic signals to improve generalization to unseen manipulation types, achieving a 2.1% average AUC improvement on the FakeAVCeleb dataset.
Detecting deepfakes, in a world where even reality is suspect
A look at the growing challenge of detecting AI-generated deepfakes as synthetic media becomes indistinguishable from reality, and the emerging tools and strategies to verify authentic content.
Harnessing CLIP and DINO: An Uncertainty-Aware Cascaded Fusion Network for Generalizable Deepfake Image Detection
This paper proposes UCF-Net, a cascaded fusion network that combines CLIP and DINO for improved deepfake image detection with better cross-domain generalization.
This light-powered AI can spot deepfakes with nearly 98% accuracy
UCLA researchers unveiled a hybrid optical-neural processor that detects deepfake videos with ~97.79% accuracy while analyzing 15-18 video streams in parallel, using light propagation to replace energy-hungry digital decoding. The system achieves 99.86% sensitivity and can be deepened with passive diffractive layers for higher accuracy at minimal energy cost.
Toward Calibrated, Fair, and accurate Deepfake Detection
Introduces Face-Fairness (FF), a plug-and-play framework for bias mitigation in deepfake detection, featuring Face-Feature Tuning (FFT) as the first demographic label-free fairness method that improves group accuracy and reduces performance gaps across demographics.