Analyzing Reasoning Shifts in Audio Deepfake Detection under Adversarial Attacks: The Reasoning Tax versus Shield Bifurcation

TL;DR

This work reframes audio deepfake detection as a reasoning-enabled task by integrating Chain-of-Thought into Audio Language Models (ALMs) and introduces a three-tier forensic audit (acoustic perception, cognitive coherence, cognitive dissonance) to study robustness under adversarial attacks. It reveals a bifurcation where reasoning acts as a shield for acoustically grounded models and as a tax for others, with linguistic attacks enabling confident yet incorrect justification. A silent-alarm mechanism via cognitive dissonance is shown to flag potential manipulation even when the final decision is compromised. The study offers a diagnostic framework for evaluating forensic audio reasoning, highlights attack-specific failure modes, and outlines limitations and avenues for defense-oriented future work.

Abstract

Audio Language Models (ALMs) offer a promising shift towards explainable audio deepfake detections (ADDs), moving beyond \textit{black-box} classifiers by providing some level of transparency into their predictions via reasoning traces. This necessitates a new class of model robustness analysis: robustness of the predictive reasoning under adversarial attacks, which goes beyond existing paradigm that mainly focuses on the shifts of the final predictions (e.g., fake v.s. real). To analyze such reasoning shifts, we introduce a forensic auditing framework to evaluate the robustness of ALMs' reasoning under adversarial attacks in three inter-connected dimensions: acoustic perception, cognitive coherence, and cognitive dissonance. Our systematic analysis reveals that explicit reasoning does not universally enhance robustness. Instead, we observe a bifurcation: for models exhibiting robust acoustic perception, reasoning acts as a defensive \textit{``shield''}, protecting them from adversarial attacks. However, for others, it imposes a performance \textit{``tax''}, particularly under linguistic attacks which reduce cognitive coherence and increase attack success rate. Crucially, even when classification fails, high cognitive dissonance can serve as a \textit{silent alarm}, flagging potential manipulation. Overall, this work provides a critical evaluation of the role of reasoning in forensic audio deepfake analysis and its vulnerabilities.

Figures (8)

• Figure 1: The proposed three-tier forensic audit framework: acoustic perception, cognitive coherence, and dissonance to analyze reasoning robustness of ALMs.
• Figure 2: Perception scores ($\Phi_{Perc}$) across six forensic dimensions, comparing the baseline sensitivity of general-purpose ALMs (dashed lines) against models fine-tuned for audio reasoning (solid lines).
• Figure 3: The Coherence-Dissonance Trade-off.
• Figure 4: Mapping the Reasoning Landscape: (left, A) The coherence landscape ($\Phi_{Coh}$ vs. ASR) and (right, B) The dissonance landscape ($\Psi_{Diss}$ vs. ASR).
• Figure 5: Reasoning as a Shield vs. Tax. Black lines are error bars indicating 95% Confidence Interval.