Successful information propagation from source to destination in Vehicular Adhoc Network (VANET) can be hampered by the presence of neighbouring attacker nodes causing unwanted packet dropping. Potential attackers change their behaviour over time and remain undetected due to the ad-hoc nature of VANET. Capturing the dynamic attacker behaviour and updating the corresponding neighbourhood information without compromising the quality of service requirements is an ongoing challenge. This work proposes a Reinforcement Learning (RL) based neighbour selection framework for VANET with an adaptive trust management system to capture the behavioural changes of potential attackers and to dynamically update the neighbourhood information. In contrast to existing works, we consider trust and link-life time in unison as neighbour selection criteria to achieve trustworthy communication. Our adaptive trust model takes into account the social relationship, time and confidence in trust observation to avoid four types of attackers. To update the neighbourhood information, our framework sets the learning rate of the RL agent according to the velocities of the neighbour nodes to improve the model's adaptability to network topology changes. Results demonstrate that our method can take less number of hops to the destination for large network sizes while can response is up to 54 percent faster compared to a baseline method. Also, the proposed model can outperform the other baseline method by reducing the packet dropping rate up to 57 percent caused by the attacker.

翻译：车载自组织网络（VANET）中，从源节点到目的节点的信息成功传播可能因邻近攻击节点导致的数据包非必要丢弃而受阻。潜在攻击者会随时间改变行为，且由于VANET的自组织特性而难以被检测。如何在满足服务质量要求的前提下捕捉攻击者动态行为并更新相应邻域信息，始终是一项持续挑战。本文提出一种基于强化学习（RL）的VANET邻居选择框架，通过自适应信任管理系统捕捉潜在攻击者的行为变化，并动态更新邻域信息。区别于现有研究，我们将信任度和链路生存时间统一作为邻居选择准则，以实现可信通信。自适应信任模型综合考虑社会关系、时间因素及信任观测置信度，可抵御四类攻击者。为更新邻域信息，该框架根据邻居节点速度调整RL智能体学习率，提升模型对网络拓扑变化的适应性。结果表明，当网络规模较大时，本方法到达目的节点的跳数更少，响应速度相比基线方法提升高达54%。同时，所提模型能将攻击者引起的数据包丢弃率降低57%，优于另一种基线方法。