Explainability by Design: Structured Kolmogorov-Arnold Networks over Probabilistic Attributes for Speech Deepfake Source Tracing

Authors: Hoang H. Pham, Manasi Chhibber, Tomi H. Kinnunen

Published: 2026-08-20 16:11:30+00:00

Comment: 26 pages, submitted to Elsevier Computer Speech & Language journal

AI Summary

This paper introduces an interpretable speech deepfake source tracing framework that combines multi-task learning for probabilistic attribute extraction with a structured Kolmogorov–Arnold Network (KAN) for attack classification. The approach provides explainability by design, allowing quantification of each probabilistic feature's contribution to source tracing without post-hoc explainers. The model achieves high accuracy and reliable interpretability on the ASVspoof2019-attr-17 dataset.

Abstract

Modern speech synthesizers can produce highly realistic speech, making source tracing (i.e. identifying the generator behind a spoofed utterance) increasingly important for forensics, online content provenance, and platform accountability. Building on our prior work on transparent probabilistic attributes, which represent utterances as probability distributions over synthesizer sub-components, we extend speech deepfake source tracing with two key ingredients: multi-task training of the probabilistic attribute extractors and a structured Kolmogorov--Arnold Network (KAN) for attack classification. The probabilistic features are estimated jointly with a multi-task learning module built on a shared AASIST or SSL-AASIST countermeasure backbone. The resulting probabilistic feature embedding is classified by a structured KAN whose topology follows known attribute-to-attack relationships. This provides interpretability by construction: the architecture reflects the generative hierarchy of attacks, while KAN feature-importance scores quantify each probabilistic feature's contribution without post-hoc explainers such as SHAP. On ASVspoof2019-attr-17, the extended framework achieves balanced accuracies above 99% for all seven probabilistic feature extractors, with EERs of 0.16% to 0.07%, and 99.64% balanced accuracy with 0.11% EER for 17-class attack classification. Our revised model outperforms the earlier two-stage baselines, in addition to demonstrating reliable interpretability, with importance scores consistent with SHAP values, and stable results across batch sizes. These findings highlight the potential of structured KAN for speech deepfake source tracing that is both accurate and interpretable by design. For transparency and reproducibility, our codebase is publicly available: https://github.com/HoangHPham/KAN-Probabilistic-Deepfake-Attribution.

Key findings

The extended framework achieved balanced accuracies above 99% for all seven probabilistic feature extractors, with EERs of 0.16% to 0.07%, and 99.64% balanced accuracy with 0.11% EER for 17-class attack classification. The model outperforms previous two-stage baselines and demonstrates reliable interpretability, with importance scores consistent with SHAP values and stable results across batch sizes.

Approach

The proposed approach uses a multi-task learning (MTL) module with a shared AASIST or SSL-AASIST backbone to extract probabilistic features representing synthesizer sub-components. These features are then fed into a structured Kolmogorov–Arnold Network (KAN) whose topology is designed to reflect known attribute-to-attack relationships, enabling direct interpretability and attack classification. The overall system is trained end-to-end.

Datasets

ASVspoof2019-attr-17 (for source tracing), ASVspoof 2019 LA (for pretraining AASIST and SSL-AASIST), ASVspoof2019-attr-2 (for pretraining AASIST and SSL-AASIST).

Model(s)

AASIST, SSL-AASIST, Kolmogorov-Arnold Network (KAN)

Author countries

Finland