Source-linked AI summary
Mitigating Degradation Attacks in Cooperative Autonomous Driving via Intention Sharing: A Vehicle-in-the-Loop Study
Prakhar Gupta, Tyler Ard, Rongyao Wang, Jagruti Sahoo, Judith Mwakalonge, Ardalan Vahidi, Yunyi Jia
TL;DR
Communication degradation can undermine safe cooperative driving, while experimental evidence for intention sharing against adversarial delays remains limited. The paper evaluates intention-sharing control in a vehicle-in-the-loop testbed and finds that the intention-sharing variants avoid collisions across the tested scenarios, unlike degraded baseline status sharing.
Problem
Experimental evidence quantifying intention sharing’s resilience to adversarial network degradation remains limited, despite the safety risks posed by delayed V2X information.
Source: §I · p. 1Method
The study evaluates status-sharing and intention-sharing control strategies in a closed-loop vehicle-in-the-loop framework under DoS-induced communication delays.
Source: §I · p. 2Results
Intention-sharing variants avoid collisions across 13 scenarios for each vehicle configuration, while baseline status-sharing controllers show degraded tracking and high collision rates.
Source: §V · p. 6Takeaways & Limitations
Abstract
from arXiv · showhide
Communication delays induced by cyber attacks present a critical challenge to the safe operation of connected autonomous driving. This study investigates the use of intention sharing communication strategy to enhance the resilience of model predictive controllers under Denial-of-Service attacks. We employ a vehicle-in-the-loop testbed integrating a real drive-by-wire vehicle with a microscopic traffic simulator and vehicle-to-X communication infrastructure. We emulate Denial-of-service attacks that induce communication delays of up to two seconds. We evaluate three controller variants: baseline status-sharing, intention-sharing, and delay-aware intention-sharing control. Experimental results reveal that while baseline control suffers significant performance degradation and frequent collisions under adversarial delay, intention sharing eliminates collisions and maintains behavior near nominal levels for the tested scenarios. These findings demonstrate the practical potential of intention-sharing architectures for safeguarding connected vehicles against network-layer degradation.
I. INTRODUCTION
Communication degradation can make cooperative driving unsafe, yet experimental evidence for intention sharing against adversarial delays remains limited. This study evaluates that mitigation in a closed-loop vehicle-in-the-loop setting and reports improved resilience in the tested car-following conditions.
I. INTRODUCTION
- Connected vehicles use V2V state sharing for cooperative car following, which can outperform traditional adaptive cruise control in safety and energy efficiency.Wireless reliance also creates exposure to network degradation attacks.Source: §I · p. 1
- DoS attacks can delay or stall legitimate V2X communication, leaving stale information that risks control-model mismatch, erratic gap regulation, and rear-end collisions.The study focuses on control and communication architectures that preserve closed-loop behavior under this network degradation.Source: §I · p. 1
- Prior work leaves limited experimental evidence on intention sharing under adversarial degradation, motivating its evaluation as a controller-level mitigation with a physical vehicle and V2X hardware.The paper identifies the joint evaluation of intention sharing, DoS-induced delays, and closed-loop physical testing as its distinguishing contribution.Source: §I · p. 1
- The study uses a vehicle-in-the-loop framework combining a microscopic traffic simulator, a real drive-by-wire vehicle, and a physical V2X stack to emulate DoS-induced queuing delays.It measures speed-tracking errors, inter-vehicle distance gaps, and collision events under nominal and adversarial conditions.Source: §I · pp. 1–2
- The proposed timestamp-based alignment mechanism uses received trajectory information and vehicle intentions to compensate for adversarial delays and improve temporal consistency in the control loop.The experiments directly compare intention sharing with status sharing under adversarial delay using a full V2X stack.Source: §I · p. 2
- Intention sharing improves resilience relative to status sharing in the car-following trials under the considered adversarial network delays.The study reports that intention-sharing variants avoid collisions across 13 scenarios for each vehicle configuration.Source: §I · p. 2
A. Control Framework for Car Following
The car-following control framework uses an MPC formulation with vehicle dynamics, bounded inputs and states, and a probabilistic constraint on the gap to the preceding vehicle. Its baseline predicts the preceding vehicle’s reference trajectory from current status, while intention sharing supplies a preview trajectory that can be time-aligned.
A. Control Framework for Car Following
- The Predictive Car-Following formulation uses an MPC terminal reference from an infinite-horizon LQR solution for following the preceding vehicle with constant time headway.Its cost penalizes control input and slack variables, while the stage index spans the N-step prediction horizon.Source: A. Control Framework for Car Following · p. 2
- The vehicle model uses longitudinal position, velocity, and acceleration states, with bounded control inputs and state constraints including maximum speed and minimum standstill gap.The acceleration response to control input has a first-order lag constant, and the continuous-time model is discretized with a zero-order hold.Source: A. Control Framework for Car Following · p. 2
- The controller constrains the minimum longitudinal gap probabilistically, treating the preceding vehicle’s future motion as stochastic and loosening confidence requirements farther into the prediction horizon.The supplied passage specifies admissible constraint probability α_i in [0.50, 1.0) and a discount factor λ = 0.95.Source: A. Control Framework for Car Following · p. 2
- In baseline PCF, the ego MPC generates the preceding vehicle’s reference position and velocity from its current status using kinematics.The reference trajectory is the input to the ego MPC for car following.Source: A. Control Framework for Car Following · p. 2
- For intention sharing, the preceding vehicle’s transmitted trajectory is previewed; nominal reception aligns preview index p with prediction index i, while delay shifts it to p = i + γ.The preview index identifies the transmitted trajectory sample used by the ego controller.Source: A. Control Framework for Car Following · p. 2
B. Threat Model for Denial-of-Service
The study models DoS as resource-exhaustion attacks that delay, rather than necessarily disconnect, infrastructure-mediated vehicle communications. It evaluates timing degradation and its scope under tested attack conditions.
B. Threat Model for Denial-of-Service
- Attacks target the RSU-facing IP interface; the RSU is the shared intermediary on the infrastructure-mediated path between roadside computing and the vehicle.The adversary can flood traffic and create concurrent connections but cannot alter the controller, legitimate messages, or their timestamps.Source: B. Threat Model for Denial-of-Service · pp. 2–3
- Flooding the roadside network can create time-varying delay and jitter in authentic messages while communication remains available, leaving controllers to use stale information.The resulting latency can range from hundreds of milliseconds to multiple seconds.Source: B. Threat Model for Denial-of-Service · pp. 2–3
- End-to-end delay is decomposed into nominal propagation and processing delay, queuing and resource-contention delay, and short-term service-time variation.The attack primarily increases the queuing component and its variability.Source: B. Threat Model for Denial-of-Service · p. 3
- The study considers delay-dominant attacks with authentic timestamped messages still arriving; PC5 jamming or flooding and control-command injection are outside scope.It evaluates reduced control degradation relative to status-only sharing, not attack detection or guaranteed safety.Source: B. Threat Model for Denial-of-Service · p. 3
- 100–2000 ms round-trip delays were observed for reasonably severe attacks; more attacking clients or higher packet rates increased average and maximum delay more than smaller payload changes.Attack parameters were varied independently, with other parameters held fixed; conclusions are limited to measured end-to-end delay within tested ranges.Source: B. Threat Model for Denial-of-Service · p. 3, Figure 2 · p. 3
C. Mitigation via Intention Sharing
Intention sharing gives the follower predicted motion samples from the preceding vehicle, and delay-aware control shifts the preview to account for latency. The approach uses feed-forward broadcasts rather than iterative vehicle consensus.
C. Mitigation via Intention Sharing
- PCF-I has the preceding vehicle broadcast its predicted motion over the MPC horizon, and the follower uses transmitted acceleration, velocity, and position samples as references.The transmitted trajectory replaces baseline extrapolated kinematic estimates.Source: C. Mitigation via Intention Sharing · p. 3
- Intention sharing provides future reference samples that let the ego MPC continue using received trajectory segments during delays within a few discretization steps, instead of extrapolating outdated data.This converts reliance on message freshness into reliance on prediction accuracy.Source: C. Mitigation via Intention Sharing · p. 4
- PCF-IDA uses packet timestamps to estimate delay and shifts the received preview forward by the corresponding sampling steps, favoring larger inter-vehicle gaps under severe latency.Broadcast intentions can become inaccurate after replanning or disturbances, so performance depends on the remaining preview being sufficiently accurate.Source: C. Mitigation via Intention Sharing · p. 4
- The system broadcasts 32 preview points at 20 Hz, with 0.5 s sampling intervals for a 16 s look-ahead; tested messages add only fixed-size parsing and alignment before the unchanged MPC.The reported implementation uses under approximately 25 ms solve time and 43.52 kbit/s for 272-byte messages at 20 Hz.Source: C. Mitigation via Intention Sharing · p. 4
- Intention sharing is feed-forward and requires no iterative communication or vehicle consensus, using predictive information in standard broadcast messages.The paper presents experiments in the following section to assess its safety and stability under adversarial delay.Source: C. Mitigation via Intention Sharing · p. 4
III. EXPERIMENTAL SETUP
The vehicle-in-the-loop evaluation combines a physical drive-by-wire vehicle with a synchronized simulated traffic environment and physical V2X communications. Two longitudinal car-following configurations support controlled comparisons of three controllers under delay.
III. EXPERIMENTAL SETUP
- The testbed couples a drive-by-wire electric vehicle equipped with RTK GPS, IMU, and LiDAR to a roadside computer running a SUMO-based digital twin.The platform uses ROS/ROS2 and a synchronized environment.Source: §III · p. 4
- A Cohda MK6 C-V2X/DSRC stack links the vehicle and roadside infrastructure through OBU and RSU devices, with clocks synchronized to GPS time.The real CAV serves as the second vehicle in both the two-vehicle and three-vehicle car-following configurations.Source: §III · p. 4
- The preceding simulated vehicle follows a scaled US06 driving-cycle segment over a 230 m lane, while attack trials produced RTT delays exceeding 2 s.The two- and three-vehicle configurations hold maneuver type, communication path, and physical test environment consistent across PCF, PCF-I, and PCF-IDA.Source: §III · p. 4
IV. RESULTS AND DISCUSSION
The evaluation compares three controller variants in two- and three-vehicle car-following configurations. Fig. 3 presents speed trajectories under nominal and attack conditions.
IV. RESULTS AND DISCUSSION
- The evaluation examines three controller variants in two- and three-vehicle car-following configurations.Source: §IV · p. 4
- Fig. 3 shows speed trajectories for the real ego CAV with the PCF controller and the preceding virtual CAV under nominal and attack conditions.Source: Figure 3 · p. 5
1) Aggregate Impact:
Across tested attack scenarios, intention-sharing variants kept speed-tracking behavior closer to nominal than status-only PCF, while control-effort changes differed between variants. The relative trends persisted in two- and three-vehicle configurations.
1) Aggregate Impact:
- RMS speed-tracking error changed by −9.3% to 92.9% for PCF, 0.01% to 4.8% for PCF-I, and −14.8% to 1.3% for PCF-IDA.Fig. 4 plots percentage changes relative to each controller’s no-attack condition across 13 attack scenarios per configuration.Source: 1) Aggregate Impact · p. 5, Figure 4 · p. 5
- RMS inter-vehicle-gap changes ranged from −44.4% to −2.9% for PCF, −0.3% to 2.7% for PCF-I, and −21.9% to 0.03% for PCF-IDA.Source: 1) Aggregate Impact · p. 5
- RMS control-effort changes ranged from 12.6% to 101.6% for PCF, 2.5% to 76.0% for PCF-I, and 9.0% to 109.4% for PCF-IDA.PCF-IDA may require greater corrective effort because its delay-aware fallback favors conservative gap preservation.Source: 1) Aggregate Impact · p. 5
- Adding a second car-following interaction did not materially change relative controller trends; PCF-I uses predictive context, while PCF-IDA prioritizes conservative gaps under severe delay.PCF-IDA may consequently require greater corrective control effort.Source: 1) Aggregate Impact · p. 5
2) Vehicle Gap and Collision Analysis:
Status-only PCF experienced frequent rear-end collisions under longer delays, whereas both intention-sharing variants avoided collisions in the conducted trials. These findings are limited to tested single-preceding-vehicle car-following scenarios and are not a theoretical safety guarantee.
2) Vehicle Gap and Collision Analysis:
- Above delays of around 1 s, baseline PCF experienced frequent rear-end collisions, with seven crash trials in each configuration; PCF-I preserved enough temporal context to avoid collisions.Source: 2) Vehicle Gap and Collision Analysis · p. 5
- PCF-IDA maintained collision-free operation across conducted trials, including at the maximum tested delay of 2.0 s, using a worst-case fallback after timestamp mismatches.Minimum following distances remained close to those of PCF-I.Source: 2) Vehicle Gap and Collision Analysis · p. 6
- The results support intention sharing as a practical resilience layer under non-ideal communication conditions.Source: 2) Vehicle Gap and Collision Analysis · p. 6
- The findings cover tested single-preceding-vehicle car-following scenarios and do not establish a theoretical safety guarantee.Communication loss, delays beyond the preview horizon, inaccurate intentions, unexpected emergency maneuvers, unmodeled disturbances, lane changes, and intersections remain unevaluated.Source: 2) Vehicle Gap and Collision Analysis · p. 6
V. CONCLUSION
The experiments provide evidence that intention sharing improves resilience under DoS-induced delays: status-sharing controllers degraded and had high collision rates, while intention-sharing variants avoided collisions across 13 scenarios per vehicle configuration.
V. CONCLUSION
- Across 13 scenarios for each vehicle configuration, intention-sharing variants avoided collisions, while baseline status-sharing controllers showed degraded tracking and high collision rates under DoS-induced delays.The results indicate that status sharing alone can be insufficient for robust cooperative driving in adversarial environments.Source: §V · p. 6