DIFF-TRO: A Self-Supervised Diffusion Framework for IoT Topology Robustness Optimization

Abstract

The robustness of Internet of Things (IoT) topologies, which measures a network’s tolerance to node removals, is critical to maintaining stable network connectivity. Existing robustness optimization methods mainly rely on search-based heuristics, which suffer from path dependence and often converge to suboptimal topologies. To address this limitation, we propose DIFF-TRO, a self-supervised diffusion framework for IoT topology robustness optimization. DIFF-TRO learns robustness-related structural priors directly through controlled noise injection and denoising in a spatio-temporal latent space, thereby mitigating path dependence. To translate latent representations into feasible and robust topologies, DIFF-TRO incorporates a constraint-aware decoding strategy that explicitly enforces geometric and connectivity constraints. In addition, a closed-loop refinement strategy progressively enhances robustness without relying on labeled data. Extensive experiments demonstrate that DIFF-TRO consistently outperforms state-of-the-art heuristic and reinforcement-learning baselines across diverse IoT network configurations.