Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery
Researchers have developed Proto-LeakNet, a novel framework designed to detect and attribute synthetic human face imagery generated by diffusion models. This system identifies subtle statistical traces, or "signal-leaks," within the latent representations of these images. Proto-LeakNet integrates closed-set classification with open-set evaluation to analyze unseen generators without requiring retraining, achieving a Macro AUC of 98.13% and demonstrating robustness against post-processing techniques. AI
IMPACT This research offers a new method for detecting synthetic media, potentially improving the authenticity verification of digital imagery.