Comprehensive Layer-wise Analysis of SSL Models for Audio Deepfake Detection

TL;DR

This study provides a detailed layer-wise analysis of SSL models for audio deepfake detection, revealing that lower transformer layers are most effective and enabling reduced computational costs without sacrificing performance.

Contribution

It offers the first comprehensive layer-wise analysis of SSL models in audio deepfake detection, highlighting the importance of lower layers and demonstrating efficient model configurations.

Findings

Lower layers are most discriminative for deepfake detection.

Models maintain competitive EER scores with fewer layers.

Using only a few lower layers reduces computational costs.

Abstract

This paper conducts a comprehensive layer-wise analysis of self-supervised learning (SSL) models for audio deepfake detection across diverse contexts, including multilingual datasets (English, Chinese, Spanish), partial, song, and scene-based deepfake scenarios. By systematically evaluating the contributions of different transformer layers, we uncover critical insights into model behavior and performance. Our findings reveal that lower layers consistently provide the most discriminative features, while higher layers capture less relevant information. Notably, all models achieve competitive equal error rate (EER) scores even when employing a reduced number of layers. This indicates that we can reduce computational costs and increase the inference speed of detecting deepfakes by utilizing only a few lower layers. This work enhances our understanding of SSL models in deepfake detection, offering valuable insights applicable across varied linguistic and contextual settings. Our trained models and code are publicly available: https://github.com/Yaselley/SSL_Layerwise_Deepfake.