Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CNCA: Toward Customizable and Natural Generation of Adversarial Camouflage for Vehicle Detectors

Linye Lyu, Jiawei Zhou, Daojing He, Yu Li

TL;DR

The proposed Customizable and Natural Camouflage Attack (CNCA) method by leveraging an off-the-shelf pre-trained diffusion model can generate natural and customizable adversarial camouflage while maintaining high attack performance.

Abstract

Prior works on physical adversarial camouflage against vehicle detectors mainly focus on the effectiveness and robustness of the attack. The current most successful methods optimize 3D vehicle texture at a pixel level. However, this results in conspicuous and attention-grabbing patterns in the generated camouflage, which humans can easily identify. To address this issue, we propose a Customizable and Natural Camouflage Attack (CNCA) method by leveraging an off-the-shelf pre-trained diffusion model. By sampling the optimal texture image from the diffusion model with a user-specific text prompt, our method can generate natural and customizable adversarial camouflage while maintaining high attack performance. With extensive experiments on the digital and physical worlds and user studies, the results demonstrate that our proposed method can generate significantly more natural-looking camouflage than the state-of-the-art baselines while achieving competitive attack performance. Our code is available at \href{https://anonymous.4open.science/r/CNCA-1D54}{https://anonymous.4open.science/r/CNCA-1D54}

CNCA: Toward Customizable and Natural Generation of Adversarial Camouflage for Vehicle Detectors

TL;DR

The proposed Customizable and Natural Camouflage Attack (CNCA) method by leveraging an off-the-shelf pre-trained diffusion model can generate natural and customizable adversarial camouflage while maintaining high attack performance.

Abstract

Prior works on physical adversarial camouflage against vehicle detectors mainly focus on the effectiveness and robustness of the attack. The current most successful methods optimize 3D vehicle texture at a pixel level. However, this results in conspicuous and attention-grabbing patterns in the generated camouflage, which humans can easily identify. To address this issue, we propose a Customizable and Natural Camouflage Attack (CNCA) method by leveraging an off-the-shelf pre-trained diffusion model. By sampling the optimal texture image from the diffusion model with a user-specific text prompt, our method can generate natural and customizable adversarial camouflage while maintaining high attack performance. With extensive experiments on the digital and physical worlds and user studies, the results demonstrate that our proposed method can generate significantly more natural-looking camouflage than the state-of-the-art baselines while achieving competitive attack performance. Our code is available at \href{https://anonymous.4open.science/r/CNCA-1D54}{https://anonymous.4open.science/r/CNCA-1D54}
Paper Structure (19 sections, 5 equations, 6 figures, 5 tables)

This paper contains 19 sections, 5 equations, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Overview
  5. Realistic Neural Rendering
  6. Adversarial Texture Generation with Diffusion Model
  7. Naturalness vs Attack Performance Trade-off
  8. Experiments
  9. Experimental Settings
  10. Attack Performance Evaluation
  11. Attack in the Digital World
  12. Attack in the Physical World
  13. Customzablility and Naturalness Evaluation
  14. Customizable Camouflage Generation
  15. Naturalness Score by Subjective Evaluation
  16. ...and 4 more sections

Figures (6)

  • Figure 1: Customized and natural adversarial camouflage with various styles. (a) A car with normal texture; (b)(c)(d)(e) are the different styles of camouflage generated by our method CNCA. Their captions are user-specified input prompts.
  • Figure 2: CNCA framework for generating customizable and natural adversarial camouflage.
  • Figure 3: Reordered texture UV map to improve camouflage naturalness.
  • Figure 4: Attack comparison on different camera poses and weather parameters. "ele" denotes elevation, "azi" denotes azimuth, "dis" denotes distance, "fog" denotes fog density, and "sun" denotes sun altitude angle. Values are car AP@0.5 (%) averaged from all models.
  • Figure 5: Real-world evaluation using two scaled cars. The upper row is the normal car model, and the bottom row is the adversarial camouflaged car.
  • ...and 1 more figures