Physics and Uncertainty Fusion Framework Defends Voice Authentication Against Deepfakes and Poisoning
Voice authentication systems operating at the network edge are vulnerable to threats from synthetic voice deepfakes and the poisoning of the distributed learning control plane. A preprint available on arXiv (identifier 2512.06040) introduces a framework that tackles these challenges by merging deepfake detection based on audio physical dynamics with uncertainty-aware learning techniques. It combines modeling of physical vocal tract dynamics with self-supervised representations, which are analyzed using a Multi-Layer Perceptron (MLP) and a Bayesian ensemble for uncertainty assessment. Additionally, the authors suggest a trust-based aggregation protocol to defend against poisoning attacks in federated learning. This framework seeks to improve detection accuracy and maintain system integrity, addressing the risks posed by AI-generated voice spoofing. No implementation specifics are included.
Key facts
- Framework couples audio physical dynamics deepfake detection with uncertainty-aware edge learning
- Fuses interpretable physics features modeling vocal tract dynamics with self-supervised learning representations
- Uses a streamlined Multi-Layer Perceptron (MLP) backbone for processing
- Bayesian ensemble provides uncertainty estimates for samples
- Trust-based aggregation protocol secures the control plane against poisoning
- Addresses dual threats: deepfake synthesis attacks and control-plane poisoning
- Designed for voice authentication systems at the network edge
- Preprint identified by arXiv:2512.06040
Entities
—