WIP: Learning Adversarial Attacks on Adaptive Traffic Signal Control Systems Under Cooperative Perception

Wangzhi Li, Tianheng Zhu, Yiheng FengVehicleSec 2025 · OpenReview

TL;DR

This paper introduces a novel black-box adversarial attack framework that leverages reinforcement learning to disrupt adaptive traffic signal control systems that rely on cooperative perception from connected and automated vehicles (CAVs). The attack successfully manipulates traffic flow, leading to a significant increase in vehicle delays, highlighting a critical vulnerability in these advanced traffic management systems.

What Problem Does This Paper Address?

Modern traffic signal control systems are becoming increasingly sophisticated, integrating advanced sensing technologies and communication protocols like Vehicle-to-Everything (V2X) to improve traffic flow. Cooperative perception, where vehicles share sensor data, is a key enabler, allowing for more accurate and efficient traffic monitoring even with a low penetration rate of CAVs. However, this increased complexity and connectivity also introduce new cybersecurity vulnerabilities. Existing research on adversarial attacks on traffic signal control often relies on white-box models (requiring knowledge of the system's internal workings) or makes unrealistic assumptions about attacker capabilities (e.g., full network access). This paper addresses the gap by proposing a practical, black-box adversarial attack that targets adaptive traffic signal control systems operating within a cooperative perception framework, aiming to cause significant traffic congestion.

Methods

The core of the proposed attack is a reinforcement learning (RL) agent trained using a multi-action proximal policy optimization (multi-PPO) algorithm. This attacker agent learns to generate a "fake CAV" and its associated "detected" vehicles. The attack operates in a black-box manner, meaning it doesn't need to know the internal workings of the traffic signal control (TSC) system.

The attacker's action space is multi-dimensional, consisting of:

  • Approach Selection: Deciding which traffic approach the fake CAV will enter.
  • Scale Action: Determining the number of fake "detected" vehicles to be associated with the fake CAV, distributed across different segments of the approach.

The attacker's goal is to maximize vehicle delay, thereby disrupting the TSC system's optimization efforts. The reward function for the attacker balances the achieved traffic congestion (increased delay) with a penalty for the "cost" of the attack (the number of fake vehicles injected), encouraging stealthy attacks. The paper models this as a multi-dimensional Markov Decision Process (MMDP).

Datasets

The experiments were conducted using a simulated real-world intersection in Ann Arbor, MI, USA, implemented in the SUMO (Simulation of Urban MObility) traffic simulation platform. The paper does not introduce new datasets but rather uses a simulated environment to represent a realistic traffic scenario.

Key Insights and Ideas

  • Vulnerability of Cooperative Perception: The paper demonstrates that even with the benefits of cooperative perception, these systems are susceptible to adversarial manipulation through falsified sensor data.
  • Effectiveness of Black-Box RL Attacks: A reinforcement learning agent can effectively learn to exploit the decision-making logic of adaptive TSC systems without prior knowledge of their internal architecture, making the attack practical in real-world scenarios.
  • Targeted Congestion: The attacker strategically targets specific approaches and segments within the intersection to maximize its impact on traffic flow, demonstrating a nuanced understanding of how traffic signals operate.
  • Trade-off between Attack and Stealth: The attacker's reward function balances the desire to cause maximum disruption with the need to remain undetected, highlighting the inherent trade-offs in designing adversarial strategies.

Key Results

The proposed adversarial attack significantly increased average vehicle delay. When compared to a system without any attack (average delay of 45.6 seconds), the proposed attack model resulted in an average delay of 74.1 seconds, a 62.5% increase. This is substantially higher than a random attack, which only increased the delay by 29.3%. The attack was particularly effective in creating congestion on main street approaches, leading to extended green phases and increased delays for other traffic movements.

Limitations and Future Work

The authors acknowledge several limitations and outline future research directions:

  • Targeting I-SIG: The current work focuses on the CAVLight system. Future work will extend the attack framework to target the I-SIG (optimization-based) traffic signal control system to compare performance differences.
  • Refining the Attack Algorithm: The multi-PPO algorithm will be refined to explore more diverse attack policies and to compare the proposed method against other benchmark attack algorithms, including those that introduce communication noise or periodic blocking.
  • End-to-End Attack Model: The research aims to develop a more comprehensive attack model that considers the trajectory generation of fake vehicles within the Sensor Data Sharing Message (SDSM), rather than just their final state.
  • Scalability to Multiple Intersections: Future work will involve scaling the experiments to multiple intersections and urban networks to understand how attackers can manipulate fake SDSMs across a larger area.
  • Detection and Defense: While not the focus of this paper, potential mitigation strategies are discussed. These include designing detection models based on physical laws to identify inconsistencies in vehicle dynamics and traffic flow, and employing defense mechanisms like falling back to fixed-time control or revoking attacker credentials.

Summary updated

BibTeX
@inproceedings{li2025wip,
  author = {Wangzhi Li and Tianheng Zhu and Yiheng Feng},
  title = {{WIP: Learning Adversarial Attacks on Adaptive Traffic Signal Control Systems Under Cooperative Perception}},
  booktitle = {VehicleSec},
  year = {2025},
  url = {https://openreview.net/forum?id=UW69yFNN2A},
}
Figure