Acoustic confrontation sample generation method based on partition disturbance

Abstract

The invention provides an acoustic confrontation sample generation method based on partition disturbance, and the method comprises the steps: dividing a key region and a non-key region on an amplitude spectrogram of an audio based on the time-frequency distribution characteristics of an audio signal according to a partition scheme; different disturbance constraints are set in different areas, disturbance is generated in different areas by using an adversarial sample generation method, and the disturbance is superposed with an original amplitude spectrum to obtain an adversarial sample amplitude spectrum. When the subarea generates disturbance, the disturbance constraint of the key area is weakened to enhance the disturbance energy of the key area, the disturbance constraint of the non-key area is enhanced to weaken the disturbance energy of the non-key area, and the generated confrontation sample has the time-frequency distribution characteristic more similar to that of the original sample; a higher attack success rate can be obtained with smaller total disturbance energy.

Description

technical field [0001] The invention relates to the field of adversarial sample generation, in particular to an acoustic adversarial sample generation method based on partition disturbance. Background technique [0002] Studying adversarial attack is helpful to understand the vulnerability of deep learning models and explore how to defend against attacks, so studying acoustic adversarial attack is very important to explore how to improve the adversarial robustness of acoustic scene classification models. The adversarial sample attack technology based on deep learning has achieved good results in attacking the acoustic scene classification model, but the typical adversarial sample attack technology generates adversarial samples with a unified standard in the global sample, and in the pursuit of a high attack success rate. , there is the problem that the adversarial perturbation in the acoustic signal adversarial sample is easy to be perceived. [0003] In order to better imp...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above deficiencies in the prior art, the present invention provides an acoustic adversarial sample generation method based on partition perturbation, which solves the problem that the current adversarial sample generation method in acoustic scene classification cannot set perturbation constraints according to the time-frequency distribution characteristics of the sound signal. Technical Issues of Adversarial Perturbation

Images