New audio deepfake detection method cuts EER by 13% using dataset identity

A team led by Mingrui Liang from Johns Hopkins University has developed a novel approach to audio deepfake detection that addresses the critical problem of cross-dataset generalization. Their paper, accepted at SPSC 2026, introduces a dataset-aware framework that uses only the dataset identity—a naturally available label—as supervisory signal for both multitask learning and gradient reversal layer training. This method allows the model to learn shared features across datasets while suppressing dataset-specific biases, overcoming limitations of prior approaches like data augmentation, auxiliary factor training, and Mixture-of-Experts (MoE), which often require predefined coverage, hard-to-obtain annotations, or high complexity.

The researchers evaluated their method on the 2025 Speech DeepFake Arena benchmark protocol, reporting results in terms of Equal Error Rate (EER). Compared to a baseline system, the multitask learning (MT) approach reduced Average EER by 13.14% relatively, while the gradient reversal layer (GRL) reduced Pooled EER by 5.32% relatively. These improvements demonstrate that the framework can boost aggregate detection performance across heterogeneous evaluation datasets, offering a practical and deployment-friendly solution for real-world scenarios where audio deepfakes may come from unknown sources or synthesis methods.