Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery

Proto-LeakNet is a novel framework designed for signal-leak aware attribution in synthetic human face imagery, addressing the challenge of source attribution in deepfake detection. It employs a combination of closed-set classification and density-based open-set evaluation within the latent domain of diffusion models, utilizing a temporal attention encoder and a feature-weighted prototype head to enhance interpretability and robustness. Achieving a Macro AUC of 98.13%, Proto-LeakNet demonstrates superior performance in distinguishing between real and synthetic images, making it a valuable tool for practitioners focused on improving the reliability of authenticity verification systems.