AdvDiffuser: Generating Adversarial Safety-Critical Driving Scenarios via Guided Diffusion

Yuting Xie, Xianda Guo, Cong Wang, Kunhua Liu, Long Chen

TL;DR

AdvDiffuser is a new framework that uses guided diffusion models to generate realistic, safety-critical driving scenarios for testing autonomous vehicles (AVs). By decoupling the "realism" of driving behavior from the "adversariness" of the scenario, it creates diverse, plausible, and highly transferable test cases that can effectively challenge various AV planners.

What Problem Does This Paper Address?

Testing autonomous vehicles in the real world is difficult because safety-critical events—like near-misses or collisions—are "long-tail" events; they happen very rarely. To solve this, researchers use simulations to generate adversarial scenarios.

However, current methods face two major hurdles:

  1. Realism vs. Adversariness: Scenarios generated to cause collisions often look "robotic" or physically impossible, making them poor benchmarks for real-world safety.
  2. Transferability: Most adversarial generators are "overfitted" to a specific AV system. If you change the AV's software, the generated scenarios might no longer be challenging, requiring a slow and expensive retraining process.

Methods

AdvDiffuser tackles these issues by using a Guided Diffusion Model. The framework consists of three main components:

  • The Diffusion Model: This model is trained on massive amounts of offline driving data (like the nuScenes dataset). It learns the "latent space" of how vehicles naturally move and interact. During generation, it performs a reverse diffusion process to turn random noise into plausible multi-agent trajectories.
  • The Reward Guide: Instead of retraining the whole model for every new AV, the authors use a lightweight "guide" (based on a Deep Q-Network or DQN architecture). This guide injects a "collision reward" gradient into the diffusion process, nudging the generated trajectories to become adversarial without breaking the rules of realistic driving.
  • Architecture: It uses a Graph Neural Network (GNN) context encoder to handle vehicle interactions and an autoregressive trajectory decoder to ensure the final paths are physically feasible (using a bicycle model).

This "decoupling" is the secret sauce: the diffusion model ensures the cars act like real cars, while the reward guide ensures they try to (safely) cause a headache for the AV being tested.

Datasets

The researchers evaluated AdvDiffuser using the nuScenes dataset, a large-scale real-world driving dataset. They used 1,000 driving scenes, splitting them into 8-second segments (2 seconds of history to predict 6 seconds of future adversarial behavior).

Key Insights and Ideas

  • Decoupling is Key: By separating the generation of realistic behavior (Diffusion) from the objective of being adversarial (Guidance), the model becomes much more flexible. You only need to "warm up" the lightweight reward model to attack a new, unknown AV planner.
  • Latent Space Diffusion: The authors found that performing diffusion in a learned latent space is far superior to diffusing directly on raw coordinate trajectories. Raw trajectory diffusion often leads to "average" behaviors that lack the diversity of real driving.
  • Few-Shot Adaptation: AdvDiffuser can adapt to a brand-new AV planner in under 100 "warm-up" steps, significantly outperforming traditional reinforcement learning methods that require thousands of interactions.
  • Efficiency: Unlike optimization-based baselines (like AdvSim or Strive) which require slow, iterative re-planning, AdvDiffuser can generate scenarios at over 10Hz, making it suitable for large-scale automated testing.

Key Results

  • Realism: AdvDiffuser achieved significantly lower Jensen-Shannon Divergence (JSD) scores for velocity and acceleration compared to baselines, meaning its generated traffic looks much more like real-world data.
  • Safety-Critical Performance: While maintaining realism, it achieved a collision rate of ~11% against a rule-based planner, comparable to or better than specialized adversarial models that produce much less realistic behavior.
  • Transferability: In cross-planner tests (e.g., training against a rule-based planner and testing against a neural-network-based SimNet planner), AdvDiffuser maintained stable and effective attack performance, whereas other methods saw significant performance drops.

Limitations and Future Work

The authors acknowledge a few areas for improvement:

  • Environmental Noise: The current model sometimes struggles with "parked cars" in the original dataset, which can skew collision metrics (treating a car parked on the curb as a "collision with the environment").
  • Likelihood vs. Severity: While the scenarios are adversarial, the paper notes a philosophical challenge in the field: how do we distinguish between a "useful" high-probability hazard and a "useless" physically impossible one?
  • Future Directions: The authors suggest extending this to more complex agents, such as pedestrians and cyclists, and investigating how these generated scenarios can be used to train more robust AVs, not just test them.

Summary updated

Abstract

Safety-critical scenarios are infrequent in natural driving environments but hold significant importance for the training and testing of autonomous driving systems. The prevailing approach involves generating safety-critical scenarios automatically in simulation by introducing adversarial adjustments to natural environments. These adjustments are often tailored to specific tested systems, thereby disregarding their transferability across different systems. In this paper, we propose AdvDiffuser, an adversarial framework for generating safety-critical driving scenarios through guided diffusion. By incorporating a diffusion model to capture plausible collective behaviors of background vehicles and a lightweight guide model to effectively handle adversarial scenarios, AdvDiffuser facilitates transferability. Experimental results on the nuScenes dataset demonstrate that AdvDiffuser, trained on offline driving logs, can be applied to various tested systems with minimal warm-up episode data and outperform other existing methods in terms of realism, diversity, and adversarial performance.
BibTeX
@inproceedings{xie2024advdiffuser,
  author = {Yuting Xie and Xianda Guo and Cong Wang and Kunhua Liu and Long Chen},
  title = {{AdvDiffuser: Generating Adversarial Safety-Critical Driving Scenarios via Guided Diffusion}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  url = {https://openreview.net/forum?id=zRp9yuwWLM},
}

Loading related content...

Figure