The invention discloses a deep reinforcement learning traffic signal control poisoning defense method based on strong disturbance detection and model retraining, and the method comprises the steps: firstly obtaining back door data in input data through strong disturbance, and then carrying out the recognition of the back door data, and further determining an abnormal data point; and finally, abnormal data points of abnormal data in the test process can be removed through defense of a data level, or forgetting learning is performed on the poisoning model through original training data and a reverse trigger reconstruction training set through defense of a model level, so that the poisoning model forgets abnormal behaviors caused by a trigger. According to the detection method, abnormal traffic state data are screened out firstly, only a reverse trigger needs to be searched in a backdoor data subset without calculating all input data, and finally, a model level and a data level are defended through a two-level defending method, so that abnormal behaviors caused by the backdoor trigger are eliminated, and the detection accuracy is improved. And the vehicle passing efficiency of the intersection is improved.

    本发明公开了一种基于强扰动检测与模型再训练的深度强化学习交通信号控制中毒防御方法,该方法首先利用强扰动获取输入数据中的后门数据,再对后门数据进行识别进一步确定异常数据点,最后可通过数据层面的防御将测试过程中异常数据的异常数据点进行去除,或是通过模型层面的防御将原始训练数据与反向触发器重构训练集对中毒模型进行忘却学习使得中毒模型忘却触发器引发的异常行为。本发明通过检测方法首先将异常交通状态数据筛选出来,只需在后门数据子集中寻找“反向触发器”而无需对所有输入数据进行计算,最后通过两个层面的防御方法对模型层面和数据层面进行防御,以此消除后门触发器带来的异常行为,提高交叉路口的车辆通行效率。


    Access

    Download


    Export, share and cite



    Title :

    Deep reinforcement learning traffic signal control poisoning defense method based on strong disturbance detection and model retraining


    Additional title:

    基于强扰动检测与模型再训练的深度强化学习交通信号控制中毒防御方法


    Contributors:
    XU DONGWEI (author) / WANG DA (author) / LI CHENGBIN (author)

    Publication date :

    2022-11-18


    Type of media :

    Patent


    Type of material :

    Electronic Resource


    Language :

    Chinese


    Classification :

    IPC:    H04L TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION , Übertragung digitaler Information, z.B. Telegrafieverkehr / G06K Erkennen von Daten , RECOGNITION OF DATA / G06N COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS , Rechnersysteme, basierend auf spezifischen Rechenmodellen / G08G Anlagen zur Steuerung, Regelung oder Überwachung des Verkehrs , TRAFFIC CONTROL SYSTEMS



    Deep reinforcement learning traffic signal control poisoning attack method based on Trojan horse attack

    XU DONGWEI / WANG DA / LI CHENGBIN | European Patent Office | 2022

    Free access

    Deep Reinforcement Learning-based Traffic Signal Control

    Ruan, Junyun / Tang, Jinzhuo / Gao, Ge et al. | IEEE | 2023


    Traffic signal control method based on deep reinforcement learning

    LIU DUANYANG / SHEN SI / SHEN GUOJIANG et al. | European Patent Office | 2021

    Free access

    Adaptive Traffic Signal Control Model on Intersections Based on Deep Reinforcement Learning

    Duowei Li / Jianping Wu / Ming Xu et al. | DOAJ | 2020

    Free access

    Deep reinforcement learning traffic signal control method based on attention mechanism

    WU JIANGUANG / ZHOU SHUYA / HOU XIANGDONG et al. | European Patent Office | 2024

    Free access