Analyzing Reasoning Shifts in Audio Deepfake Detection under Adversarial Attacks: The Reasoning Tax versus Shield Bifurcation
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.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The main contribution of this paper is the introduction of a forensic auditing framework that evaluates the robustness of reasoning in Audio Language Models under adversarial attacks, revealing critical insights into the interplay between reasoning and model performance. This comprehensive analysis not only advances the understanding of audio deepfake detection but also sets a precedent for future research in model interpretability and robustness.
Methodology
The paper introduces a novel forensic auditing framework to analyze the robustness of Audio Language Models (ALMs) under adversarial attacks, focusing on three dimensions: acoustic perception, cognitive coherence, and cognitive dissonance. This approach is innovative as it shifts the focus from traditional binary classification to understanding the reasoning process behind model predictions, providing a more nuanced view of model performance. The methodology effectively combines theoretical insights with practical auditing techniques, although it relies heavily on the specific models chosen for analysis.
Results & evaluation
The experiments are well-structured, utilizing a variety of adversarial attack strategies to assess model performance across different conditions. The use of the ASVSpoof 2019 dataset is appropriate, and the results are presented clearly, demonstrating the varying impacts of reasoning on model robustness. However, the paper could benefit from a broader range of datasets to validate the findings across different contexts.
Reproducibility
The paper provides detailed implementation information, including model fine-tuning techniques and adversarial attack frameworks. However, the absence of a publicly available code repository or demo limits reproducibility. The iterative refinement process for dataset generation is described, but without access to the code or data, independent verification is challenging.
Limitations
The study is limited to English-language datasets, and its findings may not generalize to multilingual contexts. Additionally, the focus on mid-sized ALMs restricts the applicability of the results to larger models that may exhibit different behaviors. The paper also acknowledges that it does not propose specific training objectives to mitigate identified vulnerabilities, which could be a critical area for future research.
Broader impact
This research has significant implications for the field of audio deepfake detection, particularly in enhancing the interpretability and trustworthiness of AI systems in high-stakes environments. By providing a framework for understanding reasoning shifts under adversarial conditions, the work contributes to the ongoing discourse on AI accountability and transparency. The main contribution of this paper is the introduction of a forensic auditing framework that evaluates the robustness of reasoning in Audio Language Models under adversarial attacks, revealing critical insights into the interplay between reasoning and model performance. This comprehensive analysis not only advances the understanding of audio deepfake detection but also sets a precedent for future research in model interpretability and robustness.
Author of this paper? Send a correction.