Securing Voice Authentication Applications Against Targeted Data Poisoning

Alireza Mohammadi, Keshav Sood, Asef Nazari, Dhananjay Thiruvady

Abstract: Deep neural network-based voice authentication systems are promising but remain vulnerable to targeted data poisoning attacks where adversaries substitute legitimate user utterances to gain unauthorized access. To address this, we propose a novel defense framework that integrates a regularized convolutional neural network with a K-nearest neighbors classifier, enhanced with stratified cross-validation and class weighting to counteract data imbalance inherent in such attacks. Evaluated on real-world datasets under realistic attack scenarios, our framework demonstrates significant robustness. It achieves accurate authentication even when as little as 5% of the training data is poisoned, outperforming existing state-of-the-art methods.

External IDs:doi:10.1109/mic.2025.3588068