Why Fake ? Unveiling the Semantic Vocabulary of Deepfake Detectors

Authors: Vazgken Vanian, Alexandros Doumanoglou, Dimitris Zarpalas

Published: 2026-07-08 09:56:06+00:00

Comment: Accepted at CVPRW 2026

AI Summary

This research investigates the decision-making processes of black-box deepfake detectors using Encoding-Decoding Direction Pairs (EDDP) from post-hoc Explainable AI. The method uncovers the detector's semantic vocabulary of 'real' and 'fake' features, enabling global model understanding, precise concept localization, and counterfactual analysis. This provides a deeper, more transparent comprehension of deepfake detection strategies than conventional binary classification methods.

Abstract

Deepfake (DF) technology poses a significant threat to information integrity, driving the need for robust detection methods. Most DF detectors only consider predicting a binary label for whether the input is real or fake, lacking the justification required for real-world applications like legal proceedings. Explainable DF Detection has emerged to address this limitation, but existing techniques frequently fall short by either relying on human annotations for precise artifact localization or generating superficially plausible textual explanations without grounding. This work investigates the use of post-hoc explainable AI (XAI) to analyze the decision-making process of state-of-the-art black-box DF detectors. Specifically, we employ Encoding-Decoding Direction Pairs (EDDP), a technique suitable for uncovering the concept space of DF detectors (their semantic vocabulary) as well as the mechanism for writing and reading concept information to and from internal representations. Our analysis reveals previously hidden real and fake features learned implicitly during detector training, offering nuanced explanations unattainable through conventional methods. This enables global model understanding, spatially aware concept localization, and counterfactual what-if analysis, all contributing to a deeper comprehension of DF detection strategies.

Key findings

The study successfully identified a semantic vocabulary of internal concepts, demonstrating their faithful integration into the model's decision process through concept cloning and misclassification correction. Counterfactual 'what-if' analysis causally validated the influence of these concepts on model predictions, enabling transparent and steerable deepfake detection. The identified concepts are optimized for the specific model architecture and dataset, suggesting a need for separate training for different model-dataset combinations.

Approach

The authors employ Encoding-Decoding Direction Pairs (EDDP) to analyze the internal representations of black-box deepfake detectors. EDDP identifies linear combinations of concept embeddings (semantic vocabulary) within the detector's latent space, allowing for the creation of Concept Detection Maps (CPMs) and Concept Contribution Maps (CCMs) to localize and quantify the influence of 'real' and 'fake' concepts on predictions. This post-hoc analysis reveals implicitly learned features without modifying the detector's architecture.

Datasets

FaceForensics++ (FF++)

Model(s)

Xception

Author countries

Greece

← Previous