Robust Workflow Generation via Adversarial Learning for Audio Deepfake Detection

Authors: Xiang Li, Pin-Yu Chen, Wenqi Wei

Published: 2026-09-17 11:14:58+00:00

AI Summary

The paper proposes ROGUE, a framework for robust audio deepfake detection that addresses the generalization issues of existing methods under real-world perturbations. ROGUE dynamically constructs robust detection workflows by orchestrating multiple detection tools through a dual-agent adversarial learning paradigm, where a perturbation agent generates challenging audio corruptions and a policy agent learns adaptive tool selection.

Abstract

The rapid advancement of speech synthesis and voice conversion technologies has made audio deepfakes increasingly realistic, posing serious security risks in practical applications. While existing detection methods achieve strong performance under controlled conditions, they often fail to generalize under real-world perturbations and corruptions. In this paper, we propose ROGUE, a framework that dynamically constructs robust detection workflows by orchestrating multiple detection tools. ROGUE formulates workflow generation as a sequential decision-making problem and introduces a dual-agent paradigm, where a perturbation agent generates audio perturbations and a policy agent learns to select and execute detection tools under perturbed conditions. Through adversarial learning, ROGUE enables perturbation-aware tool selection, adaptive execution strategies, and improved robustness to distribution shifts. Extensive experiments across multiple datasets and real-world corruptions demonstrate that ROGUE consistently outperforms strong baselines in both robustness and generalization. Our results highlight the effectiveness of adversarially optimized workflow generation for building reliable audio deepfake detection systems in real-world deployment settings.

Key findings

ROGUE consistently outperforms strong baselines in both robustness to diverse audio perturbations and cross-dataset generalization, achieving significant improvements especially under challenging real-world corruptions like codec distortions. The adversarial learning approach enables perturbation-aware tool selection and adaptive execution strategies, leading to a more reliable audio deepfake detection system compared to static or non-adversarial methods.

Approach

ROGUE frames workflow generation as a sequential decision-making problem using a dual-agent system. A perturbation agent generates challenging audio corruptions by sampling from a learnable distribution of perturbation tools, while a policy agent, implemented as an LLM-based controller, adaptively selects and executes detection tools based on intermediate observations. These agents are trained adversarially to optimize for robustness and generalization.

Datasets

WaveFake, LJSpeech, SONAR, ASVspoof2019, ASVspoof2021 LA, CodecFake, Fake-or-Real, LibriSeVoc, In-the-Wild, DFADD

Model(s)

HuBERT, Wav2Vec2-BERT, DF Arena 1B V1, DF Arena 500M V1, GPT-5, Claude Sonnet 4.6, Gemini 3 Pro (as backbone for workflow generator), CQCC-GMM, LFCC-LCNN, Spectrogram-ResNet, AASIST, RawGAT-ST

Author countries

USA

← Previous