Digital Face Synthesis: Safeguarding against Threats
COM3 Level 2
SR21, COM3 02-60
Abstract:
Digital media forensics is advanced in this thesis through an exploration of face synthesis, its societal impact, and detection challenges. An attack scheme is proposed, demonstrating that subtle perturbations in synthesized face images can be used to deceive black-box face recognition systems, thereby exposing inherent vulnerabilities. The societal impact of synthesized media, particularly deepfakes, is further analyzed by examining audience perceptions and factors influencing social media sharing. Additionally, a novel deepfake detection method is introduced, leveraging Vision and Timeseries Transformers to detect deepfakes at both frame and video levels. A benchmark dataset is developed to validate this approach, featuring nuanced transitions between real and fake frames. Comprehensive empirical analysis of modern deepfake detection methods is conducted, highlighting their limitations against advanced generation techniques. Overall, these findings emphasize the pressing need for detection models to evolve rapidly, bridging the critical gap with increasingly sophisticated deepfake technology for enhanced digital media integrity.