Why Fake ? Unveiling the Semantic Vocabulary of Deepfake Detectors

Research Overview

Deepfake detection models typically function as black boxes, providing a binary real/fake label without explaining the underlying evidence. While existing explainable AI (XAI) methods attempt to address this, they often rely on coarse saliency maps or generate textual justifications that lack grounding in the model's actual decision-making process. This paper introduces a post-hoc approach using Encoding-Decoding Direction Pairs (EDDP) to uncover the 'semantic vocabulary' of deepfake detectors, allowing researchers to interpret the internal latent space of these models without requiring architectural modifications or retraining.

The authors apply EDDP to the 12th residual block of a pre-trained Xception detector. By assuming a linear representation hypothesis, they decompose the model's internal representations into a set of concept embeddings. This allows them to identify specific directions in the latent space that correspond to human-interpretable features. They validate these concepts using three methods: sensitivity analysis (RCAV) to measure influence on the final logit, semantic mapping via Intersection over Union (IoU) with facial masks, and qualitative inspection of high-activation image patches.

The study successfully identifies 16 distinct concepts, some of which strongly correlate with 'real' classifications (e.g., specific nose or eye features) while others act as primary indicators for 'fake' content (e.g., artifacts around the mouth or neck). The authors demonstrate that these concepts are not merely incidental; by cloning concept coefficients onto different base representations, they can reliably manipulate the model's output. Furthermore, they show that targeted interventions on these coefficients can correct misclassified samples with high success, confirming that these concepts are the primary drivers of the detector's logic.