Physics-Inspired Attribution for Cyber-Physical IoT Systems.

This paper introduces a novel framework for interpretable AI, drawing inspiration from statistical mechanics, to provide structural attribution for complex cyber-physical IoT systems. Unlike traditional explainability methods that focus on correlations or causal methods requiring explicit directed graphs, this approach models variable dependencies using an undirected, energy-based representation. This is particularly valuable in large-scale, hybrid systems where recovering a full directed causal structure is often impractical due to feedback loops and partial observability. The framework enables rigorous, dependency-aware attribution by analyzing how changes in the system's energy landscape reflect the influence of individual components. It also supports reasoning about perturbation effects across both continuous and discrete interactions within hybrid systems, offering reliable explanations for abnormal behaviors. This method avoids the need to recover a directed causal graph, simplifying the explanation process for intricate systems. Empirical evaluations on an industrial IoT testbed, involving hybrid continuous and discrete variables, demonstrated that the proposed framework achieves higher attribution accuracy, improved robustness, and better scalability compared to state-of-the-art graph-based approaches. While not intended to fully reconstruct the system's generative dynamics, these dependency-aware explanations are highly valuable for human interpretation and for enhancing downstream predictive and diagnostic tasks, with applications extending beyond industrial IoT security to other high-dimensional cyber-physical and socio-technical systems.