Technical Summary: From Graphs to Gradients: Physics-Inspired Structural Attribution for Cyber-Physical IoT Systems and Beyond

1. Problem Statement

Interpretable Artificial Intelligence (AI) is critical for high-risk domains like Cyber-Physical Systems (CPS) and Industrial Internet of Things (IIoT), where understanding the root causes of system behaviors is essential for safety and diagnostics. Traditional explainability methods often rely on correlational associations or Structural Causal Models (SCMs) that require explicit directed causal graphs. However, recovering such directed graphs is frequently impractical in large-scale, hybrid CPS characterized by:

• Feedback loops and partial observability: Making causal directionality difficult to identify from observational data.
• Hybrid variable types: Mixing continuous (e.g., sensor readings) and discrete (e.g., actuator states) variables.
• Scalability issues: Existing graph-based explainability tools (e.g., GNNExplainer) struggle with computational complexity and robustness as system dimensionality increases.
• Model misspecification: Sensitivity to distributional shifts and adversarial conditions.

The paper argues that while causal inference is desirable, the strict requirement for recovering a directed causal structure limits applicability in complex, real-world industrial settings.

2. Methodology

The authors propose a novel energy-based attribution framework inspired by statistical mechanics. Instead of modeling directed causal chains, the system is represented as an undirected graphical model defined by an energy landscape.

Core Concepts

• Energy Function ($E(x)$): The system state $x$ is modeled using a Boltzmann distribution $p(x) = \frac{1}{Z} \exp(-\beta E(x))$.
  • Low energy values correspond to configurations consistent with normal system behavior.
  • High energy values correspond to abnormal, adversarial, or deviant states.
  • The energy function is decomposed into unary potentials $\phi_i(x_i)$ (intrinsic constraints) and pairwise potentials $\phi_{ij}(x_i, x_j)$ (interactions), allowing the model to handle hybrid continuous and discrete domains.
• Attribution Mechanisms: The framework quantifies the influence of variables through three complementary lenses:
  1. Local Sensitivity (First-Order): Computes the magnitude of the gradient of the energy function with respect to a variable ($\Delta_i = |\frac{\partial E(x)}{\partial x_i}|$). This measures how sensitive the system's energy state is to local perturbations.
  2. Global Attribution (Free Energy & Entropy): Uses concepts of free energy ($F(x_i)$) and conditional entropy to assess how fixing a variable $x_i$ reshapes the global probability distribution. This captures variables that consistently drive the system toward stable or unstable regimes across the full distribution, not just at a single point.
  3. Second-Order Curvature: Analyzes the Hessian matrix of the energy function to measure the stability of influences and interaction effects between variables ($\gamma_{ij}$). This is applied selectively to top-ranked variables to balance computational cost and interpretability.

Computational Pipeline

The framework employs Monte Carlo estimation (specifically Metropolis-Hastings) to approximate conditional expectations and entropies, which are intractable in high-dimensional hybrid spaces. The algorithm proceeds in three phases:

1. Compute local gradient sensitivities for all variables.
2. Estimate global free energy and entropy contributions via sampling.
3. Refine explanations for the top-$k$ influential variables using second-order curvature scores.

3. Key Contributions

The paper outlines three primary contributions:

1. Formulation of an Energy-Based Framework: A dependency-aware attribution method grounded in undirected graphical representations and statistical mechanics. It provides interpretable explanations for hybrid cyber-physical systems without requiring the recovery of a directed causal graph.
2. Empirical Evaluation on SWaT: A thorough evaluation on the Secure Water Treatment (SWaT) industrial testbed. The study demonstrates the framework's ability to accurately identify dominant contributing variables and their downstream effects during realistic cyber-physical attacks (e.g., sensor spoofing leading to tank overflow).
3. Superior Scalability and Accuracy: A comparative analysis showing that the proposed method achieves higher explanatory accuracy and significantly better computational scalability than state-of-the-art graph-based methods like GNNExplainer, particularly as system size increases.

4. Experimental Results

The framework was tested on the SWaT dataset, which includes heterogeneous sensor measurements, actuator states, and control signals.

• Attribution Accuracy:
  • Precision@k: The energy-based method achieved near-perfect Precision@1 as system size increased, correctly identifying the root cause (the compromised sensor) in the top rank. In contrast, GNNExplainer's Precision@1 remained close to zero across all system sizes.
  • AUC-PR: The proposed method maintained consistently higher Area Under the Precision-Recall Curve (AUC-PR) compared to GNNExplainer. While GNNExplainer's performance degraded sharply with increasing system dimensionality, the energy-based method remained robust.
• Robustness to Perturbations:
  • When subjected to additive Gaussian noise, the energy-based framework maintained a high robustness score (stability of attribution scores).
  • GNNExplainer exhibited rapid deterioration in robustness and higher variability, likely due to its reliance on optimizing instance-specific masks over noisy data.
• Scalability:
  • Runtime: The proposed framework demonstrated linear scalability $O(nMd)$ (where $n$ is variables, $M$ is samples, $d$ is graph degree) in sparse graphs.
  • Comparison: GNNExplainer exhibited super-quadratic scaling $O(Kn^2)$, leading to significantly longer runtimes as the number of variables grew.

5. Significance and Claims

The paper positions its contribution as a shift from model-centric explainability (explaining a trained predictor) to system-centric explainability (explaining the underlying generative dynamics of the physical system).

• Principled Interpretability: By leveraging statistical mechanics, the framework offers a unified mechanism for attribution that is robust to noise and adversarial conditions, avoiding the "black box" nature of training-dependent explainers.
• Handling Hybrid Systems: The method naturally accommodates mixed continuous and discrete variables, a common challenge in IIoT that many existing methods struggle with.
• Practical Utility: The authors claim the framework enables faster and more reliable fault diagnosis in critical infrastructure, potentially reducing downtime and safety risks. It provides "dependency-aware" explanations that support human interpretation and downstream diagnostic tasks.
• Limitations and Scope: The authors explicitly state that the resulting attributions are not intended to fully recover the system's underlying generative dynamics or a directed causal graph in the Pearlian sense. Instead, they provide valuable structural explanations where causal directionality is inferred through domain knowledge and the temporal progression of incidents, rather than learned solely from observational data.

The work concludes that this approach offers a scalable, interpretable alternative for structural attribution in complex, real-world environments, extending beyond industrial IoT to other socio-technical systems.