Scalable, Efficient Optical Architecture for Muxed Deepfake Video Detection (arxiv.org)

A groundbreaking hybrid digital-analog architecture has been developed for detecting deepfake videos, addressing the pressing need for effective detection systems in the face of an increasing volume of AI-generated content. This novel approach integrates a lightweight digital front-end with a spatially multiplexed optical decoding back-end, enabling the simultaneous processing of over 15 video streams in a single optical pass. This innovation significantly enhances throughput and accuracy while reducing the energy demands typical of traditional deep learning methods. The optical-neural architecture achieved remarkable results, boasting a deepfake detection accuracy of 97.79% and demonstrating resilience against various video quality issues and adversarial attacks. Its ability to maintain high sensitivity (99.86%) and specificity (95.72%) on challenging datasets, including classical and real-world deepfakes, highlights its potential to revolutionize the AI/ML landscape for media authenticity verification. By merging optical computation with AI inference, this system exemplifies a pathway towards greater efficiency and robustness in deepfake detection, setting a new benchmark for future research and deployment in this critical area.