Exploring the Janus Face of Synthetic Images: From Privacy-secure Biometrics Applications to Deepfake Detection for Misinformation-Free Social Networks

Abstract

The rise of generative AI presents a profound duality, or a “Janus Face,” for digital society. On one hand, its ability to synthesize hyper-realistic faces offers a powerful solution to long-standing privacy and data scarcity challenges in biometric systems-a promising but underexplored application. On the other hand, this same technology is weaponized to create 'deepfakes' that fuel misinformation campaigns on Online Social Networks (OSNs), posing a significant threat to digital integrity. However, countering this threat is hampered by critical failures in existing deepfake detectors. They are often: (i) Brittle in the Wild: They prove vulnerable to the compression and post-processing artifacts introduced by OSNs. (ii) Poorly Generalizable: They fail to detect forgeries from new or unseen generative models. (iii) Computationally Inefficient: Many state-of-the-art models are too parameter-heavy for practical deployment on resource-constrained devices. This dissertation confronts this duality by addressing both sides of the coin. First, it examines the “substitutability” of synthetic face data, demonstrating that biometric classifiers (e.g., Age, Gender etc.) trained on AI-generated faces can match or even exceed the generalization performance of those trained on real face data. Second, to counter the malicious use of this technology, this dissertation develops a framework of deepfake detectors designed to be robust, generalizable, and efficient by construction. My work introduces novel, lightweight feature sets on different cues (e.g., colour cue-based Relative Chrominance Difference, Gradient features, Depth cues etc.) that are inherently resilient to OSN transformations and improve generalization to unseen forgeries. Preliminary results confirm state-of-the-art performance, achieving high accuracy in challenging real-world scenarios with a significant reduction in model complexity.