CAMOU: Learning Physical Vehicle Camouflages to Adversarially Attack Detectors in the Wild

Yang Zhang, Hassan Foroosh, Philip David, Boqing GongICLR 2019

Introduction

Researchers have developed CAMOU, a novel framework that learns physical camouflage patterns designed to hide 3D vehicles from state-of-the-art AI object detectors. By using a "clone network" to bridge the gap between a non-differentiable simulator and the detector, they created patterns that significantly reduce detection rates across different environments, vehicle types, and even different AI architectures.

The Challenge of Physical Adversarial Attacks

Most adversarial attacks—tricks used to fool AI—happen in the digital world by subtly changing individual pixels in an image. However, creating a physical adversarial attack (like a pattern painted on a car) is much harder.

To be effective in the "wild," a physical camouflage must survive a wide range of transformations: different lighting, camera angles, distances, and weather conditions. Furthermore, optimizing these patterns usually requires a "differentiable" pipeline, where you can calculate exactly how a small change in the pattern affects the AI's output. Most high-fidelity simulators used to model 3D cars are non-differentiable "black boxes," making standard optimization techniques fail.

The CAMOU Methodology

The authors propose an iterative, two-threaded approach to solve this optimization bottleneck:

  1. The Clone Network ($V_\theta$): Since the simulator (Unreal Engine 4) and the object detector (e.g., Mask R-CNN) are non-differentiable when combined, the authors train a separate neural network to "clone" their joint behavior. This clone network learns to predict the detector's score based on the camouflage pattern and the environment. Because this clone is a neural network, it is fully differentiable.
  2. Camouflage Optimization: Using the clone network as a proxy, the authors use the Expectation-over-Transformation (EoT) principle. They search for a camouflage pattern that minimizes the average detection score across a massive variety of simulated viewpoints and conditions.
  3. Iterative Refinement: They find a "good" camouflage, test it in the real simulator, and then feed those new results back into the clone network to make it even more accurate in the regions where the detector is most likely to fail.

Experimental Setup

The study relies on high-fidelity simulations rather than static image datasets:

  • Environments: Two distinct 3D scenes built in Unreal Engine 4: "DownTown" (a complex urban Manhattan-like environment) and "Landscape Mountains" (a rural, high-altitude scene).
  • Vehicles: A 3D model of a 2015 Toyota Camry XLE and a virtual SUV.
  • Detectors: The attacks were tested against Mask R-CNN (with a ResNet-101 backbone) and YOLOv3-SPP, both pre-trained on the industry-standard MS-COCO dataset.

Key Insights and Results

The core innovation of CAMOU is treating the entire rendering-and-detection pipeline as a black box and learning a substitute function. This allows the researchers to use powerful gradient-based optimization on a system that was previously "un-optimizable."

  • Resolution Sweet Spot: They tested patterns from $2 \times 2$ to $256 \times 256$ pixels. Surprisingly, $16 \times 16$ was the most effective. Higher resolutions didn't help because fine details blur together when the camera is far away, whereas the $16 \times 16$ blocks remained disruptive at various distances.
  • Targeting "Attention": By using saliency maps, the authors found that detectors focus heavily on the "upper body" of the car (roof, hood, trunk). The learned camouflage works by specifically disrupting the visual features in these high-importance areas.
  • Universal Vulnerabilities: A camouflage learned to fool Mask R-CNN also worked quite well against YOLOv3. This suggests that different object detectors, despite having different architectures, share similar structural weaknesses.

Detection Drops and Qualitative Failures

In urban environments, the learned camouflage reduced the Mask R-CNN's precision ([email protected]) from 77.57% (for standard grey cars) to just 52.17%. When the camouflage trained on the Camry was applied to a completely different vehicle (the SUV), it was even more effective, dropping detection precision by nearly 44%. Furthermore, patterns trained in the city still worked in the mountains, achieving a 47% relative drop in performance without any additional tuning.

Comparison of baseline, random camouflage, and CAMOU optimized camouflage

When the attack succeeded, the AI didn't just "miss" the car; it often confidently misclassified it as entirely unrelated objects, such as a "truck," "potted plant," "boat," "kite," or even a "cake" (as seen in the optimized examples).

Limitations and Future Work

  • Legal/Functional Constraints: The authors only applied camouflage to the "paintable" body parts. They left tires, windows, and grilles untouched. If these parts were also camouflaged, the attack would likely be even more devastating.
  • Black-Box Limitation: The current method assumes no internal knowledge of the detector. The authors suggest that "white-box" access (knowing the detector's exact weights) could lead to even more effective patterns.
  • Real-World Gap: While the simulation is photo-realistic, the paper does not include tests on a physical car in the real world. Factors like paint texture, weather-induced wear, and sensor noise in real cameras remain to be fully explored.

Summary updated

BibTeX
@inproceedings{zhang2019camou,
  author = {Yang Zhang and Hassan Foroosh and Philip David and Boqing Gong},
  title = {{CAMOU: Learning Physical Vehicle Camouflages to Adversarially Attack Detectors in the Wild}},
  booktitle = {International Conference on Learning Representations},
  year = {2019},
}
Figure