Exposing Face Manipulation Based on Generative Adversarial Network-Transformer and Fake Frequency Noise Traces

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• The proliferation of realistic synthetic images generated by Generative Adversarial Networks (GANs) and diffusion models poses significant misuse risks.
• Deepfakes represent a serious societal concern due to their increasing indistinguishability from authentic images.

Purpose of the Study:

• To develop a novel deepfake detection network capable of identifying highly realistic forged face images.
• To enhance the robustness and accuracy of deepfake detection against subtle and diverse forgery patterns.

Main Methods:

• Proposed a novel detection network integrating Generative Adversarial Networks (GANs) and transformer architectures.
• Incorporated frequency domain analysis and noise detection prediction modules.
• Utilized GANs for local forgery artifact capture and transformers for global dependency modeling.

Main Results:

• The proposed framework effectively captures local artifacts and models global dependencies in forged images.
• Frequency domain and noise information were leveraged to predict anomalies, enhancing detection.
• Achieved higher accuracy and robustness compared to existing deepfake detection methods on benchmark datasets.

Conclusions:

• The novel GAN and transformer-based network offers a significant advancement in deepfake detection.
• The integration of frequency and noise analysis improves the identification of subtle and complex forgeries.
• The method provides a more accurate and robust solution for combating the misuse of synthetic media.