Abstract
Critical infrastructure networks (CINs) are increasingly vulnerable to cyberattacks that can cascade across interdependent systems. While regulatory bodies require operational data to verify attack alarms and maintain situational awareness, utilities are often reluctant to share raw sensor data due to privacy and competitive concerns. This paper introduces a two-phase differential privacy (DP) framework that bridges this gap. By combining LSTM-based state-space modeling with a novel statistical alignment technique, the framework allows regulators to verify alarms with high accuracy using noisy, private data. Evaluations on the HAI and ORNL-PS datasets show that the system achieves high alignment with non-private tests while maintaining low false alarm rates and operational efficiency.
Introduction
The security of Critical Infrastructure Networks (CINs), such as power grids and water treatment plants, is paramount due to their high level of interdependency. A localized cyberattack can quickly escalate into a regional crisis. To mitigate these risks, regulatory bodies and Information Sharing and Analysis Centers (ISACs) need utilities to report anomalies and share underlying data for verification.
However, a significant barrier exists: raw sensor data can expose trade secrets, operational costs, and systemic vulnerabilities. This creates a "trust deadlock" where regulators lack the data for effective oversight, and utilities fear the risks of data disclosure. This research proposes a "trustworthy but private" pipeline for regulatory compliance using differential privacy.
Proposed Framework
The proposed framework operates across two primary layers: the Attack Surface/OT Layer and the Compliance Layer, facilitating secure communication between the Utility Stakeholder and the Regulatory Body.
1. State-Space Modeling via LSTM-NLKF
At the core of the utility's detection system is a Non-Linear Kalman Filter (NLKF) implemented using Long Short-Term Memory (LSTM) networks. The LSTMs learn the complex, non-linear dynamics of industrial control systems (ICS) to predict future sensor states. The difference between these predictions and actual sensor readings—the "residuals"—serves as the primary indicator of a cyberattack.
2. Two-Phase Differential Privacy
To protect sensitive information before disclosure, the framework employs a two-phase privacy scheme:
- Phase 1 (Covariance Protection): System noise covariance matrices are perturbed using Laplacian noise.
- Phase 2 (Residual Protection): The low-dimensional residuals are protected using Gaussian Differential Privacy (GDP).
3. Statistical Alignment and Verification
A critical innovation of this work is the theoretical mapping that calculates a "DP-equivalent" level of significance ($\hat{\alpha}$). This allows the regulator to calibrate their statistical tests (e.g., $\chi^2$ tests) on noisy data to match the Type-I error rate (false alarm rate) of the utility's original, non-private detection. Regulators can verify data through Critical Region (CR) verification or P-value (PV) compliance.
Experimental Evaluation
The framework was validated using two real-world benchmarks: the HAI Security Dataset (complex industrial process) and the ORNL Power System (ORNL-PS) Dataset (power grid disturbances).
Key Results
- Detection Accuracy: The "DP Alignment Rate"—the frequency with which the private test matches the non-private test—reached 0.95 to 0.96 during sustained attacks.
- False Alarm Mitigation: False alarm rates remained low, between 4% and 8%, preventing regulatory alarm fatigue.
- Computational Efficiency: In containerized environments, the regulator service maintained a low overhead, with CPU utilization under 11%, demonstrating suitability for real-time streams.
Discussion and Future Work
The framework's strength lies in its mathematical guarantee against linkage attacks and its ability to "re-calibrate" significance thresholds. The introduction of a dynamic calibration factor ($\mu_t$) further incentivizes sharing by improving test power during critical events without compromising the privacy budget.
However, challenges remain. Calculating the $l_2$-sensitivity of residuals in novel or evolving systems is complex. Future research will focus on adaptive DP budgeting, where privacy levels shift based on real-time threat levels, and scaling the framework to handle the composition of privacy loss across massive, multi-utility networks.
Summary updated