An Investigation of Denial of Service Attacks on Autonomous Driving Software and Hardware in Operation

TL;DR

The paper addresses DoS vulnerabilities in autonomous driving systems by experimentally evaluating ICMP flood attacks on both an AD software stack and a GNSS-RTK localization device. It employs two setups: a Raspberry Pi-based AD stack running a tightly coupled MPC and a GNSS-RTK system connected via a wireless link, measuring real-time performance under attack. The results reveal a clear dichotomy: the AD software stack exhibits modest degradation, whereas the GNSS-RTK device experiences substantial degradation including frequent sample-rate outages and data loss. These findings highlight distinct attack surfaces and underscore the need for cybersecurity hardening, robust I/O handling, and alternative data-link strategies to ensure safe, real-time autonomous operation in adversarial environments.

Abstract

This research investigates the impact of Denial of Service (DoS) attacks, specifically Internet Control Message Protocol (ICMP) flood attacks, on Autonomous Driving (AD) systems, focusing on their control modules. Two experimental setups were created: the first involved an ICMP flood attack on a Raspberry Pi running an AD software stack, and the second examined the effects of single and double ICMP flood attacks on a Global Navigation Satellite System Real-Time Kinematic (GNSS-RTK) device for high-accuracy localization of an autonomous vehicle that is available on the market. The results indicate a moderate impact of DoS attacks on the AD stack, where the increase in median computation time was marginal, suggesting a degree of resilience to these types of attacks. In contrast, the GNSS device demonstrated significant vulnerability: during DoS attacks, the sample rate dropped drastically to approximately 50% and 5% of the nominal rate for single and double attacker configurations, respectively. Additionally, the longest observed time increments were in the range of seconds during the attacks. These results underscore the vulnerability of AD systems to DoS attacks and the critical need for robust cybersecurity measures. This work provides valuable insights into the design requirements of AD software stacks and highlights that external hardware and modules can be significant attack surfaces.

Figures (8)

• Figure 1: Overview of physical setup involving the attacker, performing an ICMP Flood attack, connected via ethernet to a Raspberry Pi running an MPC motion controller module.
• Figure 2: Overview of physical setup involving the attacker, performing an ICMP Flood attack, connected via wifi to a GNSS-RTK device.
• Figure 3: AD stack computation time per time step, with horizontal bars at minimum, median, and maximum value
• Figure 4: Positions in horizontal plane during one 30s experiment, while antennas and receiver were not moving
• Figure 5: Sample-to-sample time increments in one recording (configuration 2)