Audio Physical Dynamics Inspired Deepfake Detection for Voice Authentication Systems

August 19, 2026 | Source: arXiv AI

Tags: deepfake detection, voice authentication, federated learning, Bayesian ML, audio security

A framework fusing physics-inspired vocal tract features with self-supervised learning and a Bayesian ensemble detects voice deepfakes at network edges while resisting federated learning poisoning — addressing both attack vectors that voice authentication systems face simultaneously.

Details

Voice authentication at network edges faces two distinct attack vectors: deepfake synthesis (spoofed voice inputs) and control-plane poisoning in distributed federated learning protocols. Most prior work addresses one but not both.\n\nThis paper fuses two complementary signal types: interpretable physics features that model vocal tract dynamics (capturing biometric properties difficult for synthesis models to replicate) and representations from a self-supervised learning module (capturing learned acoustic patterns). These feed into an MLP backbone followed by a Bayesian ensemble that outputs uncertainty estimates alongside binary predictions.\n\nThe uncertainty estimates serve double duty: flagging low-confidence audio samples for additional scrutiny, and enabling a trust-based aggregation protocol in federated learning that discounts updates from potentially poisoned nodes. The combined architecture targets edge deployment — lightweight enough for distributed voice authentication infrastructure while being robust against both attack surfaces.