Abstract This paper proposes a score-based likelihood ratio system for forensic identification of deepfake images, addressing challenges in digital media identification due to rapid deepfake development. Built on the FaceForensics + + dataset, the system prevents data leakage via video-level splits (training, validation, selection, calibration, and test sets). Among six candidate models, the Capsule detector demonstrates the most robust performance (AUC = 0. 983). Score distributions of real and fake samples are modeled using kernel density estimation, with optimal bandwidths selected through a two-stage search (real: 0. 004, fake: 0. 003). Extreme LR values are bounded using the empirical lower and upper bounds method (− 2. 3634 to 1. 9933), and PAV calibration is applied to optimize the calibration performance of the LR system. On the FF + + test set, the system exhibits favorable performancewith forensic practice expectations: low misleading evidence rates (RMEP = 0. 069, RMED = 0. 092), good error control (EER = 0. 0804), and reduced decision loss after calibration (the cost log-likelihood ratio from 0. 2899 to 0. 1625). Generalization tests on five unseen datasets (Celeb-DF-v1/v2, DFDCPₘethodA/B, UADFV) yield AUCs between 0. 621 and 0. 783—highest on UADFV (0. 783), stable on DFDCP, weaker on Celeb-DF. The results show that at the moment, the technique shows potential for forensic-oriented deepfake identification, but requires further validation across diverse real-world scenarios before practical forensic application.
Guo et al. (2026) studied this question.