Physics-based deepfake detector secures voice authentication with Bayesian AI

Voice authentication systems at the network edge face two converging threats: increasingly sophisticated AI-generated deepfake audio designed to spoof a speaker's voice, and control-plane poisoning attacks that corrupt federated learning protocols. In a new arXiv paper (2512.06040), researchers Alireza Mohammadi, Keshav Sood, Dhananjay Thiruvady, and Asef Nazari propose a framework that tackles both simultaneously by coupling audio physical dynamics-inspired detection with uncertainty-aware edge learning.

The framework extracts interpretable physics features that model vocal tract dynamics—how the human voice is physically produced—then fuses them with representations from a self-supervised learning module. These combined features pass through a streamlined Multi-Layer Perceptron (MLP) backbone, followed by a Bayesian ensemble that provides uncertainty estimates for each audio sample. By incorporating both physical audio characteristics and uncertainty, the model can flag deepfake audio even when the synthesis is sophisticated enough to fool conventional detectors. On the security side, the framework's trust-based aggregation protocol secures the distributed training control plane against poisoning attacks, ensuring that maliciously injected updates don't corrupt the voice authentication model during federated learning.