An Intervention-Based Framework for Shortcut Diagnosis in Spoofing Countermeasures

Authors: Santiago Rubio, Pilar Bello, Dayana Ribas, Antonio Miguel, Eduardo Lleida, Alfonso Ortega

Published: 2026-07-03 09:42:31+00:00

Comment: Accepted at Odyssey 2026: The Speaker and Language Recognition Workshop

AI Summary

This paper introduces an intervention-based framework, grounded in a directed graphical model, to diagnose shortcut learning in deepfake audio detection systems. It formally distinguishes confound-driven shortcut dependencies from legitimate domain shift by applying controlled acoustic perturbations. The evaluation on ASVspoof datasets reveals that non-speech intervals are a dominant shortcut exploited by models.

Abstract

While deepfake audio detection systems achieve high performance in controlled benchmarks, their reliability often diminishes in the wild. Prior work shows that dataset-specific artifacts contribute to this gap. Yet, systematic tools to identify which acoustic properties a model exploits as shortcuts remain limited. We propose an intervention-based diagnostic framework, grounded in a directed graphical model, that formally distinguishes confound-driven shortcut dependencies from legitimate domain shift. We operationalise this through controlled acoustic perturbations targeting non-speech structure, spectral content, and signal energy, complemented by corpus-level distributional analysis. Evaluating XLS-R-300M with RawGAT-ST across ASVspoof challenges datasets, we quantify model sensitivity to specific intervention types. Results reveal that non-speech interventions produce the largest performance shifts, confirming non-speech intervals as a dominant shortcut.

Key findings

Non-speech interventions (e.g., adding zero-padding or AWGN to non-speech regions) resulted in the largest performance shifts, unambiguously confirming non-speech intervals as a dominant shortcut. Models trained with data augmentation specifically targeting non-speech artifacts (DA models) showed significantly reduced sensitivity to these shortcuts compared to models trained with standard augmentations (RB models). Simply scaling the dataset size without addressing these confounds is insufficient to resolve shortcut dependencies.

Approach

The authors propose an intervention-based diagnostic framework using a directed graphical model to differentiate shortcut dependencies from domain shift. They operationalize this by applying controlled acoustic perturbations (targeting non-speech structure, spectral content, and signal energy) and analyzing their impact on model performance, complemented by corpus-level distributional analysis. This quantifies model sensitivity to specific intervention types.

Datasets

ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 5

Model(s)

XLS-R-300M (front-end), RawGAT-ST (classifier)

Author countries

Spain

← Previous