Score-based Likelihood Ratios for Deepfake Image Evidence

Deepfake technology's rapid advancement challenges forensic evidence evaluation. Conventional binary detection models use fixed thresholds to classify images but lack probabilistic interpretation and cannot quantify evidential strength, limiting their forensic applicability. To address this, we propose a score-based likelihood ratio framework that combines deep feature extraction with probabilistic modeling for deepfake authentication.We created real-real and real-fake image pairs, extracted 2048-dimensional features using the SRM network, and computed pairwise similarity scores. Kernel density estimation modeled the score distributions for both pair types, with bandwidth optimized via the Silverman rule and ten-fold cross-validation for robustness. This produced an LR model that converts similarity scores into quantifiable evidential strength, enabling hypothesis testing between real and deepfake images. On the test set, the framework achieved an EER of 0.0192. Applying the Pool Adjacent Violators Algorithm reduced the classification log-likelihood ratio cost from 0.0202 to 0.0128. DET, Tippett, and ECE analyses confirmed satisfactory discrimination and calibration performance on the tested dataset. We further enhanced the model using the ELUB method, which helps improve the statistical consistency of LR outputs in the experimental setting. This study presents a rigorous, computation-driven approach to quantifying deepfake evidence. By integrating deep learning with calibrated probabilistic modeling, it aims to provide reliable and interpretable evidential strength, contributing to the potential advancement of LR-based analysis in legal and forensic settings.