Deep reinforcement learning model poisoning detection method and device based on time sequence neural pathway

Abstract

The invention discloses a deep reinforcement learning model poisoning detection method and device based on a time sequence neural pathway. The method comprises the following steps: defining a time sequence neural pathway of deep reinforcement learning, constructing a time sequence neural pathway of a deep reinforcement learning model comprising a first part comprising a convolutional layer and a pooling layer and a second part comprising a full connection layer through definition according to a time sequence nerve, namely obtaining Top-c neurons of the first part through multiple times of searching, putting the Top-c neurons and all neurons of the second part into a neuron pool, and constructing a time sequence neural pathway of deep reinforcement learning according to the neuron pool; inputting the sample data into the deep reinforcement learning model, generating disturbance by using back propagation of the constructed time sequence neural pathway, and adding the disturbance to the input sample to obtain a poisoning sample; and inputting the poisoning sample into the deep reinforcement learning model, and detecting whether the deep reinforcement learning model is poisoned or not according to the decision action change of the deep reinforcement learning model.

Description

technical field [0001] The invention belongs to the field of moderate detection, and in particular relates to a deep reinforcement learning model poisoning detection method and device based on sequential neural pathways. Background technique [0002] Deep reinforcement learning (DRL) is a new research hotspot in the field of artificial intelligence. Since its inception, deep reinforcement learning methods have achieved substantial breakthroughs in many tasks that require perception of high-dimensional raw input data and decision control. DRL has been widely used in different fields, including gaming, autonomous driving, healthcare, financial transactions, robot control, cybersecurity, computer vision, and more. [0003] Artificial intelligence technology replaces human beings to make autonomous decision-making in many fields, but recent research shows that deep reinforcement learning models are vulnerable to different types of malicious attacks, and the security vulnerabili...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the problem that the current deep reinforcement learning model is easily poisoned and difficult to detect after being poisoned, the present invention provides a method and device for detecting poisoning of a deep reinforcement learning model based on a sequential neural pathway, which can be optimized by neurons on the temporal neural pathway. Approximate poisoning test samples, through the approximate poisoning test samples to detect whether the deep reinforcement learning model is poisoned

Images