The Alpha Blending Hypothesis: Compositing Shortcut in Deepfake Detection

2026-05-11

AI Summary

This work addresses the limited generalization and opaque decision mechanisms of existing deepfake detection methods. The authors propose the Alpha Blending hypothesis, arguing that current frame-level detectors primarily rely on low-level blending artifacts introduced during face synthesis rather than semantic anomalies or generator-specific fingerprints. Building on this insight, they introduce a novel training paradigm that requires no real-world forged samples: models are trained exclusively on authentic facial images augmented with self-blending augmentation (SBI). An ensemble architecture is further designed to be both sensitive and robust to such blending artifacts. Evaluated across 15 deepfake datasets released between 2019 and 2025, the proposed approach achieves state-of-the-art average cross-dataset generalization performance, with the ensemble model attaining 94.0% AUROC.

Abstract

Recent deepfake detection methods demonstrate improved cross-dataset generalization, yet the underlying mechanisms remain underexplored. We introduce the Alpha Blending Hypothesis, positing that state-of-the-art frame-based detectors primarily 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-called self-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 average cross-dataset generalization on 15 compositional deepfake datasets released 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-art AUROC of 94.0% in an ensemble configuration. The code with experiments and the trained model will be publicly released.

Problem

• deepfake detection
• alpha blending
• compositing artifacts
• cross-dataset generalization
• self-blended images

Innovation

• Alpha Blending Hypothesis
• Deepfake Detection
• Self-Blended Images
• Cross-Dataset Generalization
• Compositing Artifacts