GMSMO: Multiobjective Evolutionary Adversarial Attacks for Remote Sensing Images

Abstract

The rapid deployment of on-orbit edge computing infrastructures, exemplified by recent initiatives to establish space-based data centers, has significantly accelerated the real-time application of remote sensing (RS) intelligence. However, this advancement exposes deep neural networks (DNNs) deployed in space to severe security threats from adversarial attacks, where imperceptible perturbations can mislead model predictions. Existing attack paradigms frame adversarial generation as a single-objective optimization problem constrained within a rigid ball, which suffers from limited decision boundaries and struggles to reconcile the conflicting objectives of adversarial transferability and visual imperceptibility. To this end, we propose a novel gradient-guided multiscale multiobjective optimization (GMSMO) framework. Specifically, we reformulate the attack generation as a balancing problem between two conflicting objectives: maximizing classification error while minimizing perturbation magnitude. This approach allows us to identify a set of optimal tradeoff solutions in which no single solution can improve one objective without compromising the other. To overcome the curse of dimensionality prevalent in pixel-level evolutionary search, GMSMO integrates the global exploration capacity of evolutionary algorithms with directional gradient-based priors. The core of this framework is the proposed multiscale momentum mutation (MSMM) operator, which extracts gradient information from surrogate models at varying scales to construct momentum-based trajectories. This mechanism effectively directs the optimization toward robust adversarial regions. Extensive experiments on standard optical RS datasets demonstrate that GMSMO generates a diverse set of adversarial examples (AEs) that achieve state-of-the-art transferability against advanced DNNs while maintaining superior visual quality compared to existing single-objective attack methods.