Abstract
Accurate and robust trajectory prediction is essential for safe and efficient autonomous driving, yet recent work has shown that even state-of-the-art prediction models are highly vulnerable to inputs being mildly perturbed by adversarial attacks. Although model vulnerabilities to such attacks have been studied, work on effective countermeasures remains limited. In this work, we develop and evaluate a new defense mechanism for trajectory prediction models based on randomized smoothing -- an approach previously applied successfully in other domains. We evaluate its ability to improve model robustness through a series of experiments that test different strategies of randomized smoothing. We show that our approach can consistently improve prediction robustness of multiple base trajectory prediction models in various datasets without compromising accuracy in non-adversarial settings. Our results demonstrate that randomized smoothing offers a simple and computationally inexpensive technique for mitigating adversarial attacks in trajectory prediction.
Introduction
In the world of autonomous driving, predicting where a nearby car or pedestrian will be in five seconds is a safety-critical task. While deep learning models (like Transformers and CVAEs) have become incredibly accurate at this, they have a "glass jaw": they are highly vulnerable to adversarial attacks. An adversarial agent can slightly alter its past trajectory—in ways that are physically plausible and nearly invisible to human observers—to trick a self-driving car's prediction model into making a massive error. This paper addresses this vulnerability by adapting randomized smoothing to the specific constraints of vehicle motion.
Methodology: Randomized Smoothing
The proposed defense is based on the principle of "smoothing" the model’s decision space. Instead of making a prediction based on a single input trajectory, the system generates $N$ versions of the observed path by adding random Gaussian noise and then averages the results to produce a final, "smoothed" prediction.
The paper explores two specific ways to apply this noise:
- Position-based smoothing: Adding noise directly to the $(x, y)$ coordinates of the observed path.
- Control-based smoothing: Adding noise to the underlying control inputs (like acceleration and steering rate) and then using a kinematic model to calculate the resulting path. This produces "noisier" paths that remain physically realistic.
Experimental Setup
The study evaluates the defense across two distinct environments:
- L-GAP: A driving simulator dataset focused on left-turn gap acceptance scenarios.
- rounD: A real-world dataset captured via drones over German roundabouts, representing complex, multi-agent interactions.
- NuScenes: Used as a supplementary dataset to pre-train models.
Key Results and Analysis
The defense showed dramatic improvements in robustness across the board. Notably, randomized smoothing significantly flattens the curve of error vs. attack magnitude. Even as the adversarial attack becomes "stronger" (larger $d_{max}$), the error of the smoothed model stays relatively stable compared to the unprotected baseline.
Key quantitative findings include:
- ADAPT on L-GAP: For the strongest attacks ($d_{max} = 1.0m$), the baseline error was 2.952m. Control-based smoothing slashed this to 1.428m.
- Trajectron++ on rounD: Under a moderate attack ($d_{max} = 0.25m$), the standard model's Average Displacement Error (ADE) skyrocketed to 4.989m. With randomized smoothing, the error dropped to just 0.341m—nearly matching its performance on clean data.
- Regularization Bonus: Surprisingly, adding a defense sometimes improved performance on normal data. In the rounD dataset for ADAPT, the smoothed model achieved a lower error (0.494m) than the original unprotected model (0.537m), likely by acting as a form of regularization.
Conclusion and Future Work
Randomized smoothing is a computationally inexpensive technique that can be applied to existing, pre-trained models without the overhead of adversarial training. While effective, future work remains to derive formal "certified" robustness guarantees for the complex outputs of trajectory prediction. Additionally, the variance among the $N$ noisy predictions could potentially serve as a valuable proxy for uncertainty in autonomous motion planning.
Summary updated