Resilience Analysis of Deep Q-Learning Algorithms in Driving Simulations Against Cyberattacks
Resilience Analysis of Deep Q-Learning Algorithms in Driving Simulations Against Cyberattacks
复制标题
DOI：
10.1109/icaic53980.2022.9896968
复制
发表时间：
2022-05
期刊：
影响因子：
--
通讯作者：
Godwyll Aikins;Sagar Jagtap;Weinan Gao
中科院分区：
文献类型：
--
作者：
Godwyll Aikins;Sagar Jagtap;Weinan Gao
深度强化学习(DRL)用于完成工程中复杂的任务，如自主驾驶，这些任务通常很难用传统的基于模型的方法来实现，因此引起了研究者的关注。随着自动驾驶汽车的安全性至关重要，以及对车辆连接的依赖日益增加，必须对网络攻击的弹性进行系统研究。在本文中，我们训练一个基于深度Q-学习的智能体在CALA模拟器中自主驾驶，以适应该智能体在网络攻击中可能经历的各种场景。具体地说，我们观察代理在存在拒绝服务和欺骗攻击时的行为和性能。结果反映了使用DRL方法对网络攻击的内在弹性水平。与传统的驾驶智能体相比，深度Q-学习智能体可以学习处理不确定性和遗漏信息，而不需要显式地对此类行为建模
Deep reinforcement learning (DRL) has attracted attentions by researchers to complete complex tasks in engineering, such as autonomous driving, that are typically very difficult to achieve using traditional model-based approaches. With the safety being critical in self-driving vehicles and the increased reliance on vehicle connectivity, the resilience to cyberattacks has to be systematically studied. In this paper, we train a deep Q-learning based agent to drive autonomously in the CARLA simulator under various scenarios that the agent may experience during a cyberattack. Specifically, we observe the agent’s behavior and performance in the presence of denial-of-service and deception attacks. The results reflect an inbuilt level of resilience to cyberattacks with the DRL methods. Comparing with conventional driving agents, deep Q-learning agents can learn to deal with uncertainty and missing information without explicitly modeling such behavior