Cheating recording detecting neural network model optimization method and system

Abstract

The embodiment of the invention provides a cheating recording detecting neural network model optimization method. The cheating recording detecting neural network model optimization method comprises the steps that a cheating recording detecting neural network model is constructed based on a feature extractor, a cheating detector and a domain predictor; source domain data and target domain data areinput into the feature extractor; the output of the feature extractor is input into the cheating detector and the domain predictor, the neural network model is detected by training cheating recording,and the loss function value of the cheating detector and the loss function value of the domain predictor are lowered; and adversarial training is conducted on the feature extractor based on the lowered loss function value of the domain predictor, and thus the deep feature output to the cheating detector by the feature extractor is feature with non-change of domain and cheating detecting distinction. The embodiment of the invention further provides a cheating recording detecting neural network model optimization system. According to the embodiment, the optimized model has no ability of distinguishing domain prediction in recording attacking detecting, and the generalization performance of cross domain testing is improved.

Description

technical field [0001] The invention relates to the field of audio detection, in particular to a method and system for optimizing a neural network model for fraudulent recording detection. Background technique [0002] Due to the convenience and reliability of identity authentication, ASV (Automatic speaker verification, automatic speaker verification) has made significant progress in deep neural networks, which has led to its commercialization in call centers, telephone banking and other applications. However, the vulnerability of ASV technology makes the ASV system vulnerable to various spoofing voice attacks. [0003] Recording spoofing attack detection technology is usually used in speaker recognition systems to detect whether the input audio is a recording attack or real audio, in order to protect the ASV system from malicious spoofing attacks. The front-end features extracted from the audio are used to train the deep learning model, which has a good distinguishing eff...

AI Technical Summary

Problems solved by technology

[0006] In order to at least solve the problem in the prior art that it is often difficult to predict the actual field of spoofed recordings, and the field identified by the spoofed recording detection neural network model trained with the same training set often does not match the actual field of spoofed recordings, that is, The problem that the neural network model for spoofing recording detection is often not effective in detecting spoofing recordings in different fields

Images