Evidence Subspace Projection: Measuring How Much Evidence Explains Deepfake Detection in Self-Supervised Speech Models

2026-07-13

AI Summary

This work addresses the challenge of quantitatively interpreting the decision-making basis of self-supervised speech models in deepfake detection. To this end, the authors propose an Evidence Subspace Projection method that constructs a shared subspace of neuron activation patterns to jointly represent multiple evidence factors and authenticity labels. By introducing a projection ratio, the approach quantifies the contribution of each evidence source to the detection decision. This method represents the first effort to enable interpretable, quantitative analysis of multi-source evidence within self-supervised models. Extensive experiments across diverse datasets and model architectures validate its effectiveness, successfully reproducing known findings while uncovering novel behavioral patterns underlying model decisions.

Abstract

Self-supervised learning (SSL) models are widely used as feature extractors for state-of-the-art audio deepfake detection, but it remains unclear how to directly and quantitatively connect what SSL models capture to detection decisions. To address this gap, we propose Evidence Subspace Projection, a method that represents both evidence factors (e.g., attack category, codec, gender, transmission) and authenticity labels in a shared space constructed from SSL models' neuron activation patterns. By projecting the decision vector onto each evidence subspace, we obtain a scalar ratio that quantifies the explanatory power of each evidence type. We evaluate SSL models in raw, fine-tuned, and post-trained settings on multiple datasets. The results confirm findings from established studies, validating the proposed method, and reveal new insights into model behavior.

Problem

• deepfake detection
• self-supervised learning
• evidence explanation
• neuron activation
• audio forensics

Innovation

• Evidence Subspace Projection
• self-supervised learning
• deepfake detection
• neuron activation analysis
• explanatory power