Stabilized Multi-Objective Optimization for Camouflage Attack in Remote Sensing

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We investigate the vulnerability of multi-task maritime surveillance systems where object detection and semantic segmentation are deployed concurrently. Existing adversarial patches often fail in the satellite domain due to extreme scale variations and gradient interference between disparate architectures. To address this challenge, we propose Stabilized Multi-Objective Optimization (SMO), a framework that jointly targets Mask R-CNN and U-Net with a single adversarial patch. Our approach introduces a logarithmic-inversion loss to provide stable, bounded gradients, with various strategies to dynamically balance seven distinct loss heads. Unlike traditional patch-based attacks where orientation, geometry and scale invariance are achieved by Expectation over Transformation (EoT), we utilize a U-Net generator to synthesize hull-aligned textures, ensuring this invariance implicitly across diverse settings. Experiments on the Airbus Ship Detection, DOTA, and VisDrone datasets demonstrate that SMO reduces both AP and mIoU significantly, exposing a critical security gap in integrated maritime monitoring frameworks.