FDIP-NET: a physics-inspired deepfake image detector leveraging FDIP luminance dynamics and calibrated CNNs

Introduction Synthetic faces generated using generative models are becoming increasingly indistinguishable from real ones, making reliable detection important for trust, security, and authenticity in digital media. Methods We propose FDIP-NET, a physics-inspired framework based on Fluid-Dynamics Inspired Processing (FDIP) that fuses RGB appearance with three luminance-gradient descriptors: a divergence-like response, gradient magnitude, and a kinetic-energy-like response. The framework combines robustness-oriented training, temperature scaling, and risk-aware threshold selection. Experiments were conducted on three publicly available facial-image datasets, DFRI, DF-wild, and RFF-140K, with an additional evaluation on FaceForensics++ c23 using its official train, validation, and test splits. Results FDIP-NET achieved mean held-out test accuracies of 98.654%, 98.952%, and 99.669% on DFRI, DF-wild, and RFF-140K, respectively, and 91.00% frame-level accuracy on FaceForensics++ c23. Discussion Component-wise comparisons indicate that luminance-gradient descriptors can complement RGB features, with their contribution depending on the dataset and selected representation. Strong within-dataset performance does not establish reliable transfer to unseen dataset distributions.