An Adversarial Perturbation Generation Method for Intelligent Single Intersection Based on CW Attack

Abstract

An adversarial disturbance generation method for intelligent single intersection based on CW attack. According to the existing DQN algorithm in deep reinforcement learning, an adaptive intelligent traffic light model for controlling single intersection is trained. The degree of preference judges whether the current state has attack value, and generates adversarial samples based on the CW attack combined with the optimal gradient sorting and the discreteness of the input state (ie, the spatial distribution of the vehicle in the corresponding road section). Finally, through the smoothness of the traffic flow in SUMO, the vehicle Changes in queue length and overall waiting time test the effect of the attack. The invention reduces the probability of an attack being discovered, and the attack effect is good.

Description

technical field [0001] The invention belongs to the security research field of reinforcement learning algorithms and models of intelligent transportation, and particularly relates to a method for generating confrontation disturbances to intelligent single-intersection based on CW attack. Background technique [0002] With the rapid development of artificial intelligence technology, AI technology has been applied to more and more aspects. As an emerging technology in artificial intelligence, deep reinforcement learning has begun to receive more and more attention from scholars and technicians due to its huge potential. Reinforcement learning has achieved a series of results in machine learning, intelligent transportation, knowledge perception and so on. [0003] In the field of intelligent transportation, due to the continuous increase in the number of motor vehicles in my country in recent years, the problem of vehicle congestion has become a huge challenge for traffic mana...

AI Technical Summary

Problems solved by technology

However, adversarial attacks against traffic lights optimized using deep reinforcement learning are still an open problem and have not been fully explored and studied.

Images