DF-UDetector: An effective method towards robust deepfake detection via feature restoration

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfakes pose significant threats to visual content authenticity and facilitate misinformation.
• Existing deepfake detection methods struggle with degraded or manipulated deepfake content.
• Degradations can compromise the effectiveness of data-centric deepfake detectors.

Purpose of the Study:

• To develop a robust deepfake detection method resilient to image degradations.
• To propose a novel approach for deepfake detection by operating in the feature space.
• To enhance the performance of deepfake detection on degraded and real-world "in the wild" deepfakes.

Main Methods:

• Proposed DF-UDetector, a three-component model for degradation deepfake detection.
• Employed an image feature extractor to capture essential image characteristics.
• Introduced a feature transforming module to map degraded features to a higher quality representation.
• Utilized a discriminator to assess the quality of transformed feature maps.

Main Results:

• DF-UDetector demonstrated comparable or superior performance against state-of-the-art methods on multiple video datasets.
• The model showed a notable performance advantage in detecting "in the wild" deepfakes.
• Feature space recovery proved effective in preserving crucial artifacts for detection despite image degradations.

Conclusions:

• The proposed DF-UDetector effectively addresses the challenge of detecting degraded deepfakes.
• Operating in the feature space for artifact recovery offers a promising direction for robust deepfake detection.
• The method shows potential for real-world applications in combating deepfake misinformation.