Fighting Deepfakes by Detecting GAN DCT Anomalies.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology poses significant risks due to misuse.
• Current deepfake detection methods using deep neural networks lack generalizability and explainability.
• Generative Adversarial Network (GAN) engines leave detectable traces in generated media.

Purpose of the Study:

• To develop a novel pipeline for detecting deepfakes.
• To identify and analyze GAN Specific Frequencies (GSF) as unique fingerprints of generative architectures.
• To improve the generalizability and explainability of deepfake detection.

Main Methods:

• Utilizing Discrete Cosine Transform (DCT) to detect anomalous frequencies.
• Analyzing the distribution of AC coefficients and inferring beta statistics.
• Implementing a pipeline to detect GAN Specific Frequencies (GSF).

Main Results:

• The proposed method successfully detects GAN-engine generated data.
• Experiments demonstrate robustness against various image attacks, including JPEG compression, rotation, and scaling.
• The technique proves innovative and surpasses current state-of-the-art detection methods.

Conclusions:

• The GSF detection pipeline offers a novel and effective approach to identifying deepfakes.
• The method provides enhanced explainability compared to existing deep learning models.
• This research contributes to more reliable deepfake detection and digital forensics.