SDCA: Towards Semantic-guided Dual Camouflage for Deceiving Human Eyes and Object Detectors

Overview

This repository presents the open-source implementation of SDCA, a novel adversarial camouflage generation framework that leverages semantic features of natural textures (e.g., color distributions and contour patterns) to optimize adversarial textures for evading both biological vision systems and computer vision systems.

Source code will be released soon.

Key Contribution

Semantic-Driven Generator (SDG)

• Utilizing programmatic noise to achieve inverse modeling of semantic features.
• Leveraging semantic features to drive the generation of the initial texture.
• Providing a priori guidance for subsequent optimization tasks.

Semantic-Constrained Optimization (SCO)

• Forming a unique semantic perturbation based on a priori semantics from the initial texture.
• Constraining the optimization space of the perturbation actively.
• Maintaining semantic consistency between adversarial textures and initial textures.

Dual-Dimensional Evaluation

• Evaluating attack performance: transferability across models/scenes and robustness under different viewpoints/occlusion conditions.
• Evaluating texture naturalness via similarity metrics (SSIM, FSIM, CSI) and the camouflage object detection (COD) task.
• Achieving state-of-the-art naturalness while maintaining transferability and robustness.

Dataset

For attack performance evaluation: [CARLA Dataset]

• Collected via [CARLA simulator]
• Download link: [Google Drive]

For naturalness evaluation

1. [RSSCN7 Dataset]

• Collected the forest-class subset only.
• Official link: [URL]

2. [Unity Jungle Scene Dataset]

• Collected via [Unity Jungle Scene Asset]
• Download link: [Google Drive]

Attack performance

Digital evaluation (detected by YOLOv5)

digital_attack.mp4

Physical evaluation (detected by YOLOv5)

physical_attack.mp4

Naturalness

2D evaluation (detected by the Canny edge detection algorithm)

3D evaluation (detected by PFNet)