FUZZY RULE-BASED ASSESSMENT OF UNCERTAINTY FACTORS IN DEEPFAKE CLASSIFICATION

Abstract

This study proposes a fuzzy logic–based expert system for evaluating uncertainty factors in deepfake image and video classification. Traditional classifiers rely primarily on prediction accuracy and often fail under non-ideal visual conditions, such as poor illumination, unstable face pose, or compression artifacts. The proposed system models three critical factors—classifier accuracy, illumination quality, and face stability—using linguistic membership functions (low, medium, high) and applies Mamdani-type fuzzy inference rules to compute a final Confidence Score. Experimental analysis shows that the fuzzy decision layer significantly improves reliability, especially in borderline cases where the model accuracy alone is insufficient to make a trustworthy conclusion. By integrating scene-quality attributes, the system reduces misclassification and provides transparent decision-making logic. This framework enhances the robustness of deepfake detection and can be effectively applied in digital forensics, security inspection, media validation, and real-time monitoring applications.