An Experimental Study on Generalization in Deep Fake Image Detection

The rise of synthetic media, particularly deepfakes, introduces both promising opportunities and critical challenges. While these technologies offer innovative avenues for entertainment, creativity, and education, they also raise significant ethical and security concerns due to their potential for misuse in disinformation campaigns and cyber threats. Identifying and mitigating deepfakes is crucial to preserving trust and safeguarding against potential harm. This thesis explores methodologies for implementing and evaluating deepfake detection models, emphasizing the importance of dataset selection, preprocessing, and model optimization techniques. Through experimentation with various models, including "UCF: Uncovering Common Features for Generalizable Deepfake Detection," "Detecting Deepfakes with Self-Blended Images," "Locate and Verify: A Two-Stream Network for Improved Deepfake Detection," and "Spatial-Phase Shallow Learning: Rethinking Face Forgery Detection in Frequency Domain," the study delves into the complexities of achieving generalization and identifies avenues for future research.