Misbehaviour detection and trustworthy collaboration in vehicular communication networks

Abstract

(English) The integration of advanced wireless technologies, e.g., cellular and IEEE 802.11p, in modern vehicles enables vehicle-to-everything (V2X) communication, fostering the next-generation Internet-of-Vehicles (IoV). The rise of IoV leads to more connected vehicles on roads, capable of making informed and coordinated decisions through real-time information sharing among vehicles, communication infrastructure, pedestrians, or roadside units (RSUs). However, V2X and IoV technologies inadvertently bring unprecedented challenges involving security and privacy vulnerabilities. Security threats and attacks can emerge from both malicious outsiders and insiders in V2X communication. Detecting and containing misbehaviours, particularly those initiated by rogue insiders, present challenging yet critical tasks for ensuring road safety. Furthermore, the pervasive use of artificial intelligence and machine learning (AI/ML) tools across various aspects poses potential threats to secure V2X operations. Motivated by these challenges, this doctoral thesis focuses on enhancing the security, robustness, and trustworthiness of V2X communications by enabling efficient and effective misbehaviour detection and fostering trustworthy collaboration. Specifically, we focus on (i) achieving effective and efficient misbehaviour detection with high accuracy and minimal false alarms, leveraging diverse spatiotemporal characteristics in vehicular data, and (ii) facilitating trustworthy information sharing for collaborative misbehaviour detection, with an emphasis on generalisability and the ability to detect previously unseen and partially observable attacks.

The absence of standardised approaches to address misbehaviours calls for advanced AI/ML-based solutions capable of handling the surging volume of data, enhancing robustness and generalisability, and meeting the real-time demands of V2X applications. To this end, we propose a generic deep RL (DRL) misbehaviour detection methodology capable of dynamically improving detection through interactions and experiences by leveraging various spatiotemporal behaviours present in the ambient vehicular measurement space. The scarcity of labelled vehicular data exacerbates the effective training of AI/ML-based models. Motivated by this challenge, we propose an ensemble learning framework for misbehaviour detection, coupled with unsupervised learning and a DRL model. This enables the detection of attacks from unlabelled vehicular data, facilitating the generalisation and detection of new and unseen attack variants. Additionally, addressing adversarial attacks poses a significant challenge, requiring enhanced solutions to make AI/ML-based misbehaviour detection more resilient against such threats. Towards this, we introduce and evaluate a tailored DRL approach designed to protect V2X communication systems against adversarial attacks, as well as mitigate issues stemming from inappropriate formatting of input training data due to vehicular sensor malfunctions or reading errors. By implementing data poisoning adversarial attacks, we demonstrate the resilience of the DRL-based misbehaviour detection approach even under severe conditions of sophisticated adversarial manipulation.

Building upon the proposed DRL-based misbehaviour detection approach, we introduce a novel scheme for collaborative misbehaviour detection. This scheme involves deploying a DRL-based misbehaviour detection model in an RSU at the network edge. It leverages transfer learning principles to share the knowledge learned about misbehaviours at the source RSUs with the target RSU, enabling the reuse of relevant expertise for collaborative misbehaviour detection. Considering data poisoning attacks aimed at influencing misbehavior detection, we implement selective knowledge transfer from trustworthy RSUs to avoid adversarial interference. We introduce a semantic relatedness metric to quantify each RSU's trust level for collaborative misbehavior detection.

(Català) L'auge de l'Internet dels vehicles incrementa els vehicles connectats, que prenen decisions mitjançant l'intercanvi d'informació en temps real amb altres vehicles o unitats de carretera (RSU). Aquestes tecnologies milloren la seguretat, l'eficiència i la sostenibilitat. Però, comporten desafiaments de seguretat i privacitat. Les amenaces poden venir de persones malintencionades o d'interns en la comunicació vehicle a tot (V2X), que poden transmetre informació falsificada o errònia, posant en perill la seguretat.

Aquesta tesi doctoral se centra a millorar la seguretat, robustesa i confiança de les comunicacions V2X, permetent la detecció eficient de conductes inapropiades, augmentant la confiança. Els objectius són: (i) aconseguir una detecció precisa de males conductes amb mínimes falses alarmes, aprofitant característiques vehicular espaciotemporals, i (ii) facilitar l'intercanvi d'informació de confiança per a la detecció col·laborativa de conductes indegudes, capaç de detectar nous atacs i atacs parcialment observables.

La manca d'enfocaments estandarditzats i les característiques úniques dels sistemes V2X plantegen desafiaments per a les solucions existents basades en dades. La tesi proposa l'ús de l'aprenentatge per reforç (RL) per detectar males conductes en xarxes V2X. Es presenta una metodologia de detecció de males conductes basada en RL profunda (DRL) que millora la detecció.

La falta de dades etiquetades dificulta l'entrenament de models de detecció basats en intel·ligència artificial/aprenentatge automàtic. Per això, es proposa un esquema combinat d'aprenentatge no supervisat i DRL, permetent la detecció d'atacs amb dades no etiquetades. També es presenta un enfocament DRL personalitzat per protegir contraatacs adversaris i gestionar problemes derivats de dades d'entrenament inadequades. Finalment, es proposa un esquema de detecció col·laborativa de males conductes basat en DRL, utilitzant RSUs per compartir coneixements de manera segura i eficient.