Forensic Lens: Deepfake Detection Through Micro-Level Facial Blood-Flow Signals

Abstract

Deepfake technology, propelled by recent advances in deep learning and most notably by generative adversarial networks (GANs), has evolved far beyond its early applications in entertainment. What began as a tool for playful visual augmentation has now emerged as a substantive challenge to digital privacy, information integrity, and public trust. As synthetic media approaches near-photorealistic fidelity, traditional detection strategies—whether based on conspicuous visual artifacts or computationally intensive convolutional neural networks—are increasingly strained. In practice, these methods often reveal shortcomings in scalability, exhibit sensitivity to dataset bias, and demand prohibitive computational resources, making them difficult to deploy in real-world scenarios. To address these limitations, this study introduces Forensic Lens, a lightweight deepfake detection framework that shifts focus from appearance-centric analysis to physiological consistency. The approach leverages remote photoplethysmography (rPPG) signals, capturing imperceptible facial color fluctuations induced by cardiovascular activity. These signals are then embedded within a similarity graph, enabling semi-supervised label propagation across both annotated and unannotated samples. By grounding detection in intrinsic physiological cues rather than purely visual patterns, the framework improves generalization while reducing reliance on large, exhaustively labeled datasets. Extensive experiments conducted on the Celeb-DF v2 benchmark demonstrate that Forensic Lens achieves an accuracy of 90%, comparable to contemporary CNN-based detectors, yet attained with markedly lower computational overhead. Beyond quantitative performance, the model offers interpretability and resilience against compression artifacts and noise—qualities often overlooked but essential in forensic practice. These characteristics make the framework particularly well-suited for deployment on resource-constrained platforms, including mobile devices and browser-based monitoring tools, where efficiency and reliability are paramount.