AD2: Analysis and Detection of Adversarial Threats in Visual Perception for End-to-End Autonomous Driving Systems

Ishan Sahu, Somnath Hazra, Somak Aditya, Soumyajit DeyWACV 2026

TL;DR

This paper investigates the surprising vulnerability of end-to-end autonomous driving (AD) systems to adversarial attacks targeting their visual perception. They demonstrate that even subtle, intermittent attacks can cause catastrophic failures, leading them to propose AD², a novel, lightweight detector that leverages spatial-temporal consistency to identify these threats in real-time.

What Problem Does This Paper Address?

Autonomous driving (AD) systems promise to revolutionize transportation, but their safety and reliability remain paramount concerns. While AD systems have made significant progress, their robustness against adversarial attacks, particularly in real-world, closed-loop scenarios, is largely underexplored. Existing research often focuses on individual AD modules (like object detection) or assumes white-box access to the system, which is unrealistic. This paper addresses this gap by analyzing the impact of various black-box adversarial attacks on state-of-the-art end-to-end AD systems, highlighting their fragility and proposing a practical solution for detecting such threats.

Methods

The authors first conduct a comprehensive analysis of adversarial threats on end-to-end AD systems using the CARLA simulator. They simulate three distinct attack vectors targeting the visual perception pipeline:

  • Poltergeist: An acoustic attack that simulates image blur.
  • Steal Now, Attack Later (SNAL): A digital attack that injects "ghost objects" into images.
  • Electromagnetic Signal Injection Attack (ESIA): An attack that introduces colored strips and distortions into images.

These attacks are applied under a restricted black-box setting, meaning the adversary cannot access the target model's parameters but can optimize perturbations using surrogate models or predefined manipulations. The attacks can be continuous or intermittent.

To counter these threats, they propose AD² (Adversarial Attack Detection for Autonomous Driving), a lightweight, real-time detection framework. AD² is a deep neural network that integrates:

  1. Convolutional Feature Extraction: Using a ResNet backbone to extract features from input camera images.
  2. Transformer-based Consistency Modules:
    • Spatial Transformer: Models consistency across overlapping fields of view from multiple cameras to detect spatial irregularities.
  3. Temporal Transformer: Models coherence across consecutive time-steps to detect temporal inconsistencies.
  4. Classification Head: A multi-class classifier that predicts whether the input is benign or one of the three attack types.

AD² is designed to operate externally, without requiring modifications to the AD agent or access to its internal latent outputs, making it suitable for practical deployment.

Datasets

The experiments are conducted within the CARLA simulator [8]. For the attack evaluation, they use the RouteScenario0 of the Longest6 benchmark. For training and evaluating the AD² detector, they construct a dataset using continuous multi-camera image streams from CARLA, with pairs of frames collected at 1-second intervals. This dataset includes combinations of clean and attacked frames, leading to four classes: benign, Poltergeist, SNAL, and ESIA. The training and testing splits are kept disjoint by using different routes.

Key Insights and Ideas

  • End-to-end AD systems are surprisingly fragile: Even intermittent, black-box adversarial attacks on the visual perception pipeline can lead to catastrophic driving failures, significantly degrading performance. This highlights a critical safety concern for current AD technology.
  • Intermittent attacks are potent: The study reveals that attacks occurring as infrequently as twice per second can destabilize driving, demonstrating that continuous attacks are not always necessary to cause harm.
  • Spatial-temporal consistency is key for detection: Adversarial perturbations often violate the natural consistency of visual input, both across different camera views (spatial) and over time (temporal). AD² leverages this insight by using transformers to model these consistencies.
  • Transformer's global attention is advantageous: Unlike CNNs with fixed receptive fields, transformers can model long-range dependencies, making them effective at detecting localized perturbations that might otherwise be missed.
  • Practical detection is feasible: AD² is designed to be lightweight and operate externally, making it a practical solution for real-time deployment in AD systems without requiring invasive modifications.

Key Results

  • Severe performance degradation: Under attacks like Poltergeist, SNAL, and ESIA, the driving scores of state-of-the-art agents (Transfuser and Interfuser) dropped by up to 99% in the worst-case scenarios.
  • Intermittent attacks cause failures: Even with attack intervals of d=11 (meaning attacks every 11 frames), significant performance degradation was observed.
  • AD² achieves high detection accuracy: The proposed AD² detector demonstrates consistently superior performance across all attack classes, achieving nearly perfect True Positive Rate (TPR) and False Positive Rate (FPR) trade-offs. It outperforms baseline detection methods.
  • Computational efficiency: AD² is significantly more computationally efficient than existing methods like CyberDet, with a 20x reduction in parameter count and 1.6x faster inference, crucial for real-time applications.
  • Agent differences: The Interfuser agent, with its built-in safety controller, showed slightly better robustness compared to Transfuser, though it sometimes drove too cautiously, leading to timeouts.

Limitations and Future Work

  • Black-box assumption: The current evaluation assumes attackers do not know about the detection mechanism. Developing adaptive attacks that can bypass AD² is a crucial next step.
  • Real-world validation: The experiments were conducted in simulation. Real-world evaluation and deployment are necessary to confirm the effectiveness and feasibility of AD².
  • Attack mitigation: While AD² detects attacks, the paper suggests future work on implementing actual mitigation strategies, such as switching to safe backup controllers upon detection.
  • Alternative architectures: Exploring other neural architectures for detection and further refining the transformer-based consistency modeling are potential avenues for future research.

Summary updated

BibTeX
@inproceedings{sahu2026ad,
  author = {Ishan Sahu and Somnath Hazra and Somak Aditya and Soumyajit Dey},
  title = {{AD2: Analysis and Detection of Adversarial Threats in Visual Perception for End-to-End Autonomous Driving Systems}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2026},
}
Figure