Addressing the poor performance of detection on early-stage adversarial samples in speaker recognition tasks, we propose a detection algorithm based on locality sensitive hashing (LSH-GPD) for adversarial sample of early-stage during the generation process in this paper. Firstly, a locality-sensitive hashing algorithm tailored to adversarial audio samples is designed, achieving rapid retrieval of similar speech samples and addressing the challenge of similarity search in high-dimensional speech data. Subsequently, the FastDTW algorithm is utilized to further group similar speech samples, eliminating hash collisions. Finally, adversarial samples are identified by determining whether the number of samples within each group exceeds a threshold. Adversarial samples generated on public datasets, the TIMIT, Common Voice, and Voxceleb2, are used to evaluate the LSH-GPD algorithm. Experimental results demonstrate that the LSH-GPD algorithm outperforms both the Voting and WaveGuard algorithms, the accuracy of detection for adversarial samples achieves 99.9%, and an early detection rate achieves 99.0%.
