Graphs to Gradients: New Physics-Inspired AI Explains IoT Systems Better

Interpreting why an AI system makes decisions in complex cyber-physical IoT environments has been a major challenge—especially when feedback loops and partial observability make directed causal graphs impractical. A team of five researchers (Evangelatos, Diou, Papadopoulos, Markakis, Sarigiannidis) proposes a novel alternative: instead of recovering explicit causal links, they model variable dependencies through an undirected, energy-based representation inspired by statistical mechanics. This 'structural attribution' framework analyzes how variations in the energy landscape reflect the influence of individual components, enabling rigorous dependency-aware explanations without needing a full causal structure.

The team tested their approach on an industrial IoT testbed with hybrid continuous and discrete variables, comparing it against state-of-the-art graph-based methods. Results show higher attribution accuracy, improved robustness to noise, and better scalability for large-scale systems. While the attributions don't fully recover generative dynamics, they provide valuable insights for diagnosing abnormal behaviors and supporting human interpretation. The framework is not limited to IoT security—it can also be applied to other high-dimensional cyber-physical and socio-technical systems, marking a significant step toward more reliable explainability in high-risk domains.