Structured Adversarial Camouflage via Voronoi Diagrams

The paper introduces a method for generating adversarial camouflage using Voronoi diagrams, optimizing seed-point locations within fixed, printable color palettes to create structured patterns. Evaluated against person detection metrics on COCO, the technique shows robust transferability across various detector models (YOLOv9-12) but suffers from reduced effectiveness when colors are repainted. This approach offers a parameter-efficient alternative for adversarial attacks, enhancing visual plausibility while compromising the performance of real-time detection systems, highlighting its implications for security in AI applications.