Title: SARA: Evaluating Reasoning in Audio Deepfake Detection

Abstract:

Audio Language Models (ALMs) represent a significant advancement toward explainable audio deepfake detection (ADD). By generating reasoning traces, these models move past opaque \textit{black-box} classifiers, offering greater transparency into their decision-making processes. However, the reliability of this reasoning is questionable; it may lack coherence with the model’s final predictions or, more critically, construct plausible yet misleading justifications for incorrect classifications. Furthermore, the robustness of ALM reasoning against adversarial attacks has not been thoroughly investigated, casting doubt on the practical trustworthiness of these explanatory capabilities.

To bridge this gap, we present \textbf{SARA} (\textbf{S}hift \textbf{A}nalysis of \textbf{R}easoning in \textbf{A}udio), a diagnostic framework designed to assess ALM reasoning across three key dimensions: acoustic perception, reasoning-verdict coherence, and dissonance. Our evaluation involves subjecting five open-source ALMs to both acoustic and linguistic adversarial attacks. The results indicate that acoustic attacks substantially impair reasoning-verdict coherence, resulting in an average drop of 14.20\% and often triggering internal logical contradictions. In contrast, linguistic attacks demonstrate higher success rates while preserving the coherence of the reasoning process.

Additionally, we reveal that the textual coherence of the generated reasoning traces acts as a latent signal for identifying adversarial inputs. This allows for the effective detection of perturbed audio, achieving an F1 score of 0.78, all \textit{without requiring access to the raw acoustic signal}. These insights highlight that reasoning traces retain diagnostic value even when the model’s final classification outputs are compromised.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC