Adversarial Reinforcement Learning Framework for Benchmarking Collision Avoidance Mechanisms in Autonomous Vehicles

Abstract

With the rapidly growing interest in autonomous navigation, the body of research on motion planning and collision avoidance techniques has enjoyed an accelerating rate of novel proposals and developments. However, the complexity of new techniques and their safety requirements render the bulk of current benchmarking frameworks inadequate, thus leaving the need for efficient comparison techniques unanswered. This work proposes a novel framework based on deep reinforcement learning for benchmarking the behavior of collision avoidance mechanisms under the worst-case scenario of dealing with an optimal adversarial agent, trained to drive the system into unsafe states. We describe the architecture and flow of this framework as a benchmarking solution, and demonstrate its efficacy via a practical case study of comparing the reliability of two collision avoidance mechanisms in response to intentional collision attempts.

Security and Privacy Issues in Intelligent Transportation Systems: Classification and Challenges

“…While AI safety and security research is gaining traction, it would be of interest to study the relevant aspects of this research to ITS technologies. On the other hand, recent research [84] [93] proposes that AI techniques may prove to be of significant value in automating the discovery, mitigation, and defense against security threats within the highly complex ITS.…”

“…Vulnerability assessment of ITS requires further research, in particular, there is a pressing need for a comprehensive vulnerability assessment framework. While some studies have focused on vulnerability assessment of particular components of ITS (e.g., [84] [93]), there is still no standard and comprehensive framework for analysis and quantification of vulnerabilities in the integrated system, particularly from a CAS point of view. Similarly, the bulk of the current literature focuses on implementing security measures post-development, leaving much to be done in establishing guidelines and frameworks for secure design and development.…”

Adversarial retraining attack of asynchronous advantage actor‐critic based pathfinding

“…There are many studies that explore the security problem in reinforcement learning. Behzadan and Munir et al 31 discovered that the self‐driving platooning vehicles can collide with each other when their observation data are manipulated. Drones equipped with RL techniques can be commanded to collide to a crowd or a building 32,33 …”

Deceptive Reinforcement Learning Under Adversarial Manipulations on Cost Signals

“…From (5), we know that the resulting change in Q * under malicious falsification is bounded by the change in the cost with a Lipschitz constant 1/(1 − β ). To see this, we randomly generate 100 falsifications h ∈ R 3×2 using randi(10) * rand (3,2) in Matlab. For each falsified costc = c + h, we obtain the corresponding Q-factorsQ * .…”

“…This type of adversarial behavior poses a great threat to CPS. For example, self-driving platooning vehicles can collide with each other when their observation data are manipulated [2]. Similarly, drones can be weaponized by terrorists to create chaotic and vicious situations where they are commanded to collide to a crowd or a building.…”