camouflage-attacks-vehicle-detection

name	camouflage-attacks-vehicle-detection
title	In-the-Wild Camouflage Attacks: Vehicle Detector Evasion via Conditional Image Editing
version	0.0.3
engine	skillxiv-v0.0.3-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2603.19456
keywords	["Adversarial Examples","Vehicle Detection","Camouflage Attacks","Image Editing","Object Detection Robustness"]
description	Craft stealthy vehicle appearances that fool detectors by formulating attacks as conditional image editing. Apply image-level stylization (match surroundings) and scene-level strategies (match semantic concepts), achieve 38% AP50 reduction with 85%+ human perceptual success, and transfer to black-box detectors.

In-the-Wild Camouflage Attacks on Vehicle Detection

Domain Problem: Adversarial Vulnerability in Vehicle Detection

Object detectors are critical for autonomous driving, aerial surveillance, and security systems. Yet they remain vulnerable to adversarial attacks that deceive machines while appearing normal to humans—a distinct challenge from traditional adversarial examples that produce visible artifacts.

Gap Analysis: Prior texture-overlaid attacks are obviously fake upon human inspection (e.g., random color patches). Real-world camouflage requires seamless visual integration with environment and scene semantics.

Adaptation Recipe: Conditional Image Editing as Attack

Reframe Attack as Image Editing: Instead of pixel-level adversarial perturbations, use conditional image editing (ControlNet) to manipulate vehicle appearance while preserving structure and realism.

Two-Stage Pipeline:

1. No-Box Attack: Fine-tune ControlNet to transfer visual style from reference images while preserving vehicle structure.

   • Input: Vehicle image, reference style image
   • Output: Vehicle with reference style applied
   • Goal: Fool detector without optimization loop

2. White-Box Attack: Add adversarial loss against target detector.

   • Combines ControlNet output with cross-entropy loss against detector
   • Fine-tunes diffusion model to maximize detection loss while maintaining realism

Two Stylization Strategies:

1. Image-Level Stylization: Match vehicle appearance to immediate surroundings (like chameleon color matching).

   • Extract reference patches from background around vehicle
   • Apply style transfer to vehicle region
   • Result: Vehicle color/texture matches local environment

2. Scene-Level Stylization: Match vehicle to semantic concepts in the scene (like grasshopper resembling leaves).

   • Identify semantic objects in scene (grass, trees, pavement, buildings)
   • Generate vehicle texture matching semantic concept
   • Result: Vehicle "blends into" category, not just colors

# Camouflage loss function combines three objectives:
# L = λ_structure * L_struct + λ_style * L_style + λ_adv * L_adv
#
# L_struct: L-channel (luminance) in LAB space