Phoneme-level voiceprint recognition confrontation sample construction system and method based on neural network generative model

Abstract

The invention discloses a phoneme-level voiceprint recognition confrontation sample construction system and method based on a neural network generative model. The system comprises a phoneme recognizer, a disturbance generator, a hearing suppressor, an alternative classifier, and a system optimizer, the suppressed disturbance generated by the hearing suppressor and the aligned voice generated by the phoneme recognizer are superposed to generate an adversarial sample, the adversarial sample is classified by the substitution classifier, scores are sent to the system optimizer, gradients are reversely propagated to the disturbance generator for iterative updating, and a trained phoneme disturbance generator is obtained. According to the method, the phoneme information in the voice is creatively fused to carry out disturbance construction, the phonemes in the voice are recognized and positioned by using the phoneme recognizer, and the fine-grained general adversarial disturbance is generated at the phoneme level, so that the adversarial disturbance generated in one step can be reused for any voice text input, finally, text-independent and input-independent general adversarial sample generation is realized, and the generation efficiency of the adversarial sample is greatly improved.

Description

Technical field [0001] The present invention relates to a sound pattern identification and a counterproductive sample, and Background technique [0002] As the most natural and direct communication method of human being, speech has always placed an important role in human-computer interaction and individual identification areas, and gradually become a popular biometric technology. At the same time, it is benefited from the rapid development of deep learning theory and technology, and the sound pattern recognition is applied to a variety of mature products (such as voice assistants, sound locks, etc.) to provide intelligent services for people's work. Studies have shown that 2020 global speech biological identification markets exceed 1.1 million US dollars and is expected to reach $ 3.9 million in 2026, which fully shows the broad development prospects of voice technology. However, after the bright foreground of the sound identification, the deep study of the shadow that is suscep...

AI Technical Summary

Problems solved by technology

Existing adversarial sample attacks in the voiceprint domain are divided into white-box attacks and black-box attacks: white-box attacks assume that the attacker knows the internal details of the target model, but the attack performance on unknown models is significantly reduced, so its practical significance is greatly limited. Large limitation; black-box attack assumes that the attacker has no prior knowledge about the target model, but requires a large number of query accesses to the target model to estimate the gradient, which is not feasible in practical attack scenarios

In addition, existing attack methods generate one-off adversarial samples for a single definite speech through complex iterative optimization algorithms, which have the disadvantages of deterministic input, time-consuming, and non-transferable

Images