Proteus: Automated Adversarial Robustness Testing for Audio Deepfake Detectors

2026-06-28

AI Summary

Existing audio deepfake detectors lack robustness under real-world audio transformations and are vulnerable to adversarial perturbations. This work proposes Proteus, a framework that combines breadth-first search with a Q-learning agent to efficiently explore a high-dimensional space of audio transformations—including codec compression, noise addition, and reverberation—to automatically construct adversarial example chains that effectively deceive detectors while preserving speech intelligibility and speaker identity. By systematically uncovering detector vulnerabilities through such adaptive attacks, Proteus leverages the discovered strong adversarial examples to guide targeted retraining, substantially enhancing the robustness of deepfake detection models.

Abstract

We present Proteus, a framework developed at Resemble AI for automated robustness testing of our audio deepfake detection system. Given a detector, Proteus systematically searches over sequences of everyday audio transformations (codec transcoding, additive noise, reverberation, dynamic-range compression, and VoIP simulation) to find combinations that fool the detector while preserving speech quality. We propose two complementary search strategies: (1) a breadth-first search that exhaustively maps augmentation effectiveness across the parameter space, and (2) a Q-learning agent designed to efficiently discover deeper attack chains by exploiting structural patterns in the BFS data. We report findings from continuous deployment of Proteus against our production detector, showing that specific augmentation chains can reliably flip detection verdicts while preserving speech intelligibility and speaker identity. We discuss how these findings are used to harden the detector through targeted retraining.

Problem

• audio deepfake detection
• adversarial robustness
• audio transformations
• speech quality preservation
• detector evasion

Innovation

• adversarial robustness
• audio deepfake detection
• automated testing
• reinforcement learning
• audio transformations