ED⁴: Explicit Data-Level Debiasing for Deepfake Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake detection models often fail to generalize due to intrinsic biases in training data.
• Existing methods are susceptible to content and forgery-specific biases, and a novel spatial bias favoring central image features.
• This spatial bias leads detectors to expect forgery clues at the image center, hindering performance on diverse real-world data.

Purpose of the Study:

• To introduce ED4, a novel data-level strategy to mitigate multiple biases in deepfake detection.
• To enhance the generalizability of deepfake detection models by addressing spatial and other intrinsic biases.
• To provide a unified framework for bias reduction that is model-agnostic and easily integrated.

Main Methods:

• Developed ClockMix to generate mixed facial images, preserving structure and increasing data diversity across identities, backgrounds, and forgery types.
• Proposed the Adversarial Spatial Consistency Module (AdvSCM) to prevent feature extraction biased by spatial location.
• Implemented ED4 as a plug-and-play debiasing strategy for existing deepfake detectors.

Main Results:

• ED4 effectively addresses spatial bias and improves deepfake detection generalizability.
• ClockMix significantly expands the training data distribution, exposing detectors to more varied scenarios.
• AdvSCM successfully constrains feature extraction, mitigating spatial bias.

Conclusions:

• ED4 offers a simple, effective, and unified approach to tackle biases in deepfake detection at the data level.
• The proposed methods, ClockMix and AdvSCM, demonstrably enhance detector performance and robustness.
• ED4 represents a significant advancement in creating more generalizable deepfake detection systems.