SDCA: Towards semantic-guided dual camouflage for deceiving human eyes and object detectors

Adversarial camouflage has gained widespread attention for its ability to prevent object detectors from accurately identifying physical-world targets. However, existing methods typically initialize textures with random values and exclude perturbation constraints during optimization. This approach lacks explicit guidance for generating perturbation patterns that conform to natural texture semantics (e.g., color distributions and contour structures), making the perturbations easily detected by biological vision systems. Regarding these problems, we propose Semantic-guided Dual Camouflage Attack (SDCA) from the perspective of joint optimization of natural semantics and adversarial perturbations. The core of SDCA consists of the Semantic-Driven Generator (SDG) and the Semantic-Constrained Optimization (SCO) strategy. SDG uses procedural noise to inversely model visual features of natural textures to achieve semantically-driven texture initialization. Meanwhile, SCO constrains the perturbation based on prior semantic information, preserving semantic consistency between the adversarial texture and the initial texture. Ultimately, SDCA can generate highly natural camouflage textures, achieving dual evasion of both biological vision systems and computer vision models. Experimental results on various detection tasks show that SDCA outperforms existing works in terms of naturalness while maintaining competitive robustness and transferability. The datasets and visualizations of SDCA are available at: https://github.com/Haoq1nYuan/Semantic-guided-Dual-Camouflage-Attack.