Sparse deepfake detection promotes better disentanglement

Authors: Antoine Teissier, Marie Tahon, Nicolas Dugué, Aghilas Sini

Published: 2025-10-07 09:03:39+00:00

AI Summary

This paper proposes a novel approach to enhance deepfake detection by introducing sparse latent representations in the AASIST architecture. By applying a TopK activation on the last hidden layer, the method improves detection performance, achieving an EER of 23.36% on ASVSpoof5 with 95% sparsity. Furthermore, it demonstrates that these sparse representations lead to better disentanglement of attack-related information in the latent space, thus promoting interpretability.

Abstract

Due to the rapid progress of speech synthesis, deepfake detection has become a major concern in the speech processing community. Because it is a critical task, systems must not only be efficient and robust, but also provide interpretable explanations. Among the different approaches for explainability, we focus on the interpretation of latent representations. In such paper, we focus on the last layer of embeddings of AASIST, a deepfake detection architecture. We use a TopK activation inspired by SAEs on this layer to obtain sparse representations which are used in the decision process. We demonstrate that sparse deepfake detection can improve detection performance, with an EER of 23.36% on ASVSpoof5 test set, with 95% of sparsity. We then show that these representations provide better disentanglement, using completeness and modularity metrics based on mutual information. Notably, some attacks are directly encoded in the latent space.

Key findings

The proposed sparse deepfake detection approach improved performance, achieving an EER of 23.36% on the ASVSpoof5 test set with a high sparsity level of 95%. Sparse latent representations, induced by TopK activation, demonstrated better disentanglement towards attack types, reflected by improved completeness and modularity metrics. A deeper analysis revealed that specific dimensions in the sparse latent space directly encode unique attacks, enhancing the interpretability of the model's decisions.

Approach

The authors modify the AASIST deepfake detection architecture by incorporating a TopK activation function on its last hidden layer. This activation enforces sparsity by retaining only the k highest activation values and setting others to zero, even with an oversized hidden layer. The sparse latent representations are then utilized for deepfake classification and are analyzed for their impact on both detection performance and the disentanglement of attack factors.

Datasets

ASVspoof 2024 (ASVspoof5) challenge dataset, MLS (English part)

Model(s)

AASIST, TopK activation

Author countries

France

← Previous