Data generalization method and system for replay attack detection system

Abstract

An embodiment of the present invention provides a data generalization method for a replay attack detection system. The method includes: constructing a replay attack detection system based on a feature extractor, a replay deception attack detector, a spoof category domain discriminator, and a real category domain discriminator; The domain data is input to the feature extractor, and the output of the feature extractor is respectively input to the corresponding replay deception attack detector, deception category domain discriminator, and real category domain discriminator to train the replay attack detection system; based on the real category Domain discriminator, loss function of spoofed category domain discriminator adversarial training of feature extractor. The embodiment of the present invention also provides a data generalization system for a replay attack detection system. The embodiments of the present invention improve the generalization performance of the system in out-of-set scenarios, while basically not affecting its in-set performance, so that the system as a whole has the best generalization performance.

Description

technical field [0001] The invention relates to the field of speaker recognition, in particular to a data generalization method and system for a replay attack detection system. Background technique [0002] Replay attack is one of the main forms of attack against speaker recognition systems, and the replay attack detection system is used to detect whether the input audio is a replay attack or real audio, so as to protect the speaker system. At present, most of the existing replay attack detection technologies focus on the scene in the set, such as using Light CNN (Convolutional Neural Networks, convolutional neural network) and ResNet models, which have good generalization effects in the same data set . [0003] These technologies use a strong deep neural network model to learn a deep representation with good distinguishing ability from suitable front-end features, and then classify the input audio very well (distinguish whether the audio is replaying attack voice or real v...

AI Technical Summary

Problems solved by technology

Although a replay attack detection system based on domain adversarial training has been proposed and has achieved certain performance improvements, the system still has defects: since only one domain discriminator is used, it can only be fitted from the level of the entire data set. The distribution of different categories cannot be specifically considered for fine-grained distribution alignment

This method leads to misalignment between replayed data and real data between different domains, which affects the ability of the system to distinguish replayed attacks from real audio

Images