Outlier-Aware Explainable AI for Deepfake Detection in Political Media

Narayan & Fox (2024) made strides in deepfake detection using Error Level Analysis and CNNs, yet, like most deepfake research, they focus on “average case” performance. But what about the rare, high-impact failures—especially when they occur in critical political contexts? Inspired by Ayadi et al. (2017) and Tian et al. (2024) on outlier detection, this idea proposes a dual-pathway system: a primary deepfake classifier (e.g., InceptionResnetV1), and an outlier detection module (leveraging vision-based outlier detection or autoencoders) that monitors both input data and model activations, flagging samples where the model’s confidence or feature representations deviate from the norm. Crucially, this system integrates explainable AI (like Grad-CAM as in Tian et al.) to provide interpretable visualizations whenever an outlier or misclassification occurs. This not only boosts detection of “unexpected” deepfakes (e.g., those crafted with novel manipulations) but also offers human-understandable explanations for failures—vital for journalistic and governmental response to misinformation. The novelty lies in fusing vision-based outlier detection and explainability directly into the deepfake pipeline, rather than treating them as post-hoc processes.

References:

1. Outlier detection based on data reduction in WSNs for water pipeline. A. Ayadi, Oussama Ghorbel, M. BenSaleh, A. Obeid, M. Abid (2017). International Conference on Software, Telecommunications and Computer Networks.
2. A Computer Vision Model to Identify Deepfakes to Combat Misinformation During Political Events. Aditya Narayan, Susan Fox (2024). IEEE Uttar Pradesh Section International Conference on Electrical, Computer and Electronics Engineering.
3. VOD: Vision-Based Building Energy Data Outlier Detection. Jinzhao Tian, Tianya Zhao, Zhuorui Li, Tian Li, Haipei Bie, Vivian Loftness (2024). Machine Learning and Knowledge Extraction.