Explainable Deepfake Detection: An Attention-Based Analysis of Cross-Dataset Generalization

Abstract

Most deepfake detection methods suffer from limited generalization and performance degradation under varying data quality and compression levels. In this work, we analyze the cross-dataset behavior of EfficientNet-B0-based deepfake detectors trained on Celeb-DF and FaceForensics++ (C32 and C40). Using Grad-CAM visualizations, we conduct an attentionbased analysis complemented by quantitative metrics including entropy, Gini coefficient, and center-to-peripheral attention ratio. Our results show that models trained on compression-heavy datasets rely on highly localized facial artifacts, achieving strong in-domain performance but poor cross-dataset generalization. In contrast, models trained on high-quality data exhibit more distributed attention patterns that correlate with improved robustness. These findings highlight the importance of attentionaware evaluation for developing explainable and generalizable deepfake detection systems.