Generalizable Speech Deepfake Detection via Information Bottleneck Enhanced Adversarial Alignment

View on arXiv ← Back to list

Authors: Pu Huang, Shouguang Wang, Siya Yao, Mengchu Zhou

Published: 2025-09-28 03:48:49+00:00

AI Summary

This paper addresses the challenge of distribution shifts in speech deepfake detection by proposing the Information Bottleneck enhanced Confidence-Aware Adversarial Network (IB-CAAN). IB-CAAN aims to learn robust and shared discriminative features by suppressing attack-specific artifacts and minimizing nuisance variability across domains. The method achieves state-of-the-art generalization performance across several standard and cross-dataset benchmarks.

Abstract

Neural speech synthesis techniques have enabled highly realistic speech deepfakes, posing major security risks. Speech deepfake detection is challenging due to distribution shifts across spoofing methods and variability in speakers, channels, and recording conditions. We explore learning shared discriminative features as a path to robust detection and propose Information Bottleneck enhanced Confidence-Aware Adversarial Network (IB-CAAN). Confidence-guided adversarial alignment adaptively suppresses attack-specific artifacts without erasing discriminative cues, while the information bottleneck removes nuisance variability to preserve transferable features. Experiments on ASVspoof 2019/2021, ASVspoof 5, and In-the-Wild demonstrate that IB-CAAN consistently outperforms baseline and achieves state-of-the-art performance on many benchmarks.

Key findings

IB-CAAN consistently outperformed the ERM baseline across all out-of-domain evaluations, demonstrating significant gains in generalizability, especially when using XLSR-based backbones. The proposed system achieved state-of-the-art performance on the ASVspoof 5 open condition (4.67% EER with augmentation) and attained the lowest EER on the challenging In-the-Wild dataset (4.93% EER without augmentation). Ablation studies confirmed that both the Information Bottleneck and Confidence-Aware Adversarial alignment components contribute substantially and complementarily to the improved generalization capabilities.

Approach

IB-CAAN combines a Variational Information Bottleneck (VIB) module and a Confidence-Aware Adversarial Network (CAAN). The VIB compresses irrelevant input variability (covariate shift) to preserve transferable features, while the CAAN uses classifier confidence scores to guide adversarial alignment, adaptively suppressing attack-specific artifacts (concept shift) without erasing crucial discriminative cues.

Datasets

ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 2021 DF, ASVspoof 5, In-the-Wild (ITW)

Model(s)

RawBMamba, XLSR+Linear, XLSR+MLP

Author countries

China, USA

← Previous