Deep learning model poisoning attack detection method and device based on mutual information

Abstract

The invention discloses a deep learning model poisoning attack detection method based on mutual information. The method comprises the following steps: (1) obtaining a sample set and a to-be-detected deep learning model; (2) pre-training a deep learning model; (3) generating a poisoning model pool and a trigger sample pair; and (4) judging whether the deep learning model is poisoned or not by utilizing mutual information. The invention also discloses a deep learning model poisoning attack detection device based on mutual information, which is used for implementing the method. The method has good applicability, can effectively judge whether the model is poisoned or not and find out a poisoned target class, and can obtain a better detection effect.

Description

technical field [0001] The invention relates to the technical field of poisoning detection, in particular to a mutual information-based deep learning model poisoning attack detection method and device thereof. Background technique [0002] Deep learning has gradually become a research hotspot and mainstream development direction in the field of artificial intelligence. Deep learning is a computational model composed of multiple processing layers, and a machine learning technique that learns data representations with multiple levels of abstraction. Deep learning represents the main development direction of machine learning and artificial intelligence research, and has brought revolutionary progress to the fields of machine learning and computer vision. Artificial intelligence technology has made breakthroughs in the fields of computer vision and natural language processing, ushering in a new round of explosive development of artificial intelligence. Deep learning is key to ...

AI Technical Summary

Problems solved by technology

Most existing testing techniques for autonomous driving rely on manual collection of test data for different driving conditions, which becomes unacceptably expensive as test scenarios increase

At the same time, the existing testing technologies are all based on the detection of poisoning attacks that are visible to triggers. The detection effect of feature embedding attacks that are not visible to triggers is very poor, and there are problems such as time-consuming and low efficiency in the detection process.

Images