Multi-Level Statistical Model Guidance Improves Generalization for Biometric Synthetic Face Detection

Energy-Based Models (EBMs) modify dataset entropy to claim high calibration and adversarial robustness. Previously hindered in real-world tasks by a reliance on generative methods, a Non-Generative EBM (NGEBM) has been developed. However, it's accuracy in real-world tasks is unprovenThe use of Class Activation Maps (CAMs) to guide model attention in choosing features during training has seen increasing support. Within the task of synthetic face detection the Directed Region Of Interest Diminution (DROID) has used CAM entropy to achieve high accuracy, but it's calibration and adversarial robustness have not been demonstratedBoth utilize statistical methods to modify entropy at different levels, each offering distinct benefits. In this paper we introduce a multiple loss function component model, Regional and Energy-Based Integrated Salience (REBIS), with three loss function terms building on EBM and CAM entropy approaches. REBIS brings both calibration and adversarial robustness to synthetic face detection, maintaining its state-of-the-art accuracy.