Multi-modal deep learning model vulnerability analysis method and system

Abstract

The invention provides a multi-modal deep learning model vulnerability analysis method and system, and belongs to the field of countermeasure attack and model robustness analysis. According to the method, robustness of a target model is tested by generating a series of white-box adversarial samples. The method mainly comprises the following steps: firstly, obtaining a target multi-modal deep learning model and a training data set used by the target multi-modal deep learning model; respectively extracting a visual modal feature, a text modal feature and an audio modal feature in the data set; respectively training single-mode local models by using the extracted data set features; testing the training effect of the single-mode local model to obtain adversarial sample generation weights of different modes; the visual mode and the audio mode generate disturbance by using a gradient descent method based on PGD, and the natural language text mode generates disturbance by using a synonym replacement method of gradient approximation optimization; and obtaining a vulnerability analysis result of the target model through the attack success rate of a series of generated multi-modal adversarial sample inspection models.

Description

technical field [0001] The invention relates to the field of adversarial attack and model robustness analysis, in particular to a method and system for analyzing the vulnerability of a multimodal deep learning model. Background technique [0002] With the continuous development of deep learning-related technologies and the increasing number of application scenarios in academia and industry in recent years, the demand for simultaneous analysis of multiple modal data has emerged, which has promoted the development of multi-modal machine learning. Multimodal machine learning is mainly used to find the correlation between multiple modalities and use the information of multiple modalities to make decisions at the same time. The most common ones are natural language modalities, sound signal modalities, and visual signal modalities. In the past ten years, multimodal machine learning has fully entered the era of deep learning, and has been widely used in tasks and scenarios such as ...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the vacancy of multimodal deep learning model adversarial sample generation and robustness analysis methods in the existing research, the present invention proposes an analysis method and system for the vulnerability of multimodal machine learning models

Images