Researchers have developed a framework for detecting deepfake audio and securing voice authentication systems. The system uses physics-based features to model vocal tract dynamics and combines them with self-supervised learning representations. It then uses a neural network to process the information and provide uncertainty estimates, making it more robust against advanced deepfake attacks. Additionally, the system includes a trust-based aggregation protocol to prevent control-plane poisoning in distributed federated learning protocols. --- Why it matters: This matters because voice authentication systems are vulnerable to sophisticated deepfake synthesis attacks and control-plane poisoning, which can compromise security. This framework provides a solution to these problems by detecting deepfakes more effectively and securing the control plane. Source: https://arxiv.org/abs/2512.06040