Power-Grid Controller Anomaly Detection with Enhanced Temporal Deep Learning

Zecheng He, Aswin Raghavan, Guangyuan Hu, Sek M. Chai, Ruby B. LeeTrustCom/BigDataSE 2019 · OpenReview

TL;DR

This paper introduces a novel anomaly detection system for power-grid controllers that leverages temporal deep learning and Hardware Performance Counters (HPCs). By modeling normal controller behavior and analyzing deviations in the Reconstruction Error Distribution (RED) of HPCs, the system can effectively detect zero-day cyberattacks with high accuracy and minimal latency, even without prior knowledge of attack patterns.

What Problem Does This Paper Address?

Power grids are critical infrastructure, and attacks on their control systems can have catastrophic consequences, as demonstrated by past incidents like the Ukraine power grid cyberattack. Detecting these attacks, especially "zero-day" attacks (which have no known signatures or behaviors), is extremely challenging. Traditional security measures often struggle with the unique constraints of power-grid controllers, such as limited computational resources and strict real-time requirements. Furthermore, transferring data off the controller for analysis creates a larger attack surface and introduces delays. The paper aims to address the critical need for robust, real-time anomaly detection in power-grid controllers that can identify unknown threats without relying on pre-existing attack signatures.

Methods

The core of the proposed solution is a data-driven defense system based on Reconstruction Error Distribution (RED) of Hardware Performance Counters (HPCs), enhanced by temporal deep learning.

Here's a breakdown of the methodology:

  • Data Source: Hardware Performance Counters (HPCs): Instead of relying on software-level logs or network traffic, the system monitors low-level hardware events (like instruction counts, cache misses, branch instructions) captured by HPCs. These counters are:
  • Tamper-resistant: Malware cannot easily access or modify them.
  • Low-cost and efficient: They are built into modern processors and require minimal overhead for data collection.
  • Temporal Deep Learning Model: A deep learning model, specifically Long Short-Term Memory (LSTM) networks (and explored Conditional Restricted Boltzmann Machines - CRBMs), is trained using only normal HPC readings from legitimate processes running on the controller. This model learns to predict the controller's normal temporal behavior.
  • Reconstruction Error Distribution (RED): After training, the model is used to predict future behavior based on historical HPC data. The "reconstruction error" is calculated as the squared difference between the predicted and observed HPC values. The distribution of these errors (RED) forms a profile of normal controller behavior.
  • Anomaly Detection: During online operation, the system continuously monitors HPCs, predicts behavior, and calculates the current RED. A Kolmogorov-Smirnov (KS) test is then used to statistically compare the observed RED with the baseline RED profile. A significant deviation indicates an anomaly, suggesting a potential attack.
  • Offline Training and Online Detection: The process is divided into two phases:
  • Offline Training and Profiling: Train the deep learning model on normal data and establish the baseline RED profile.
  • Online Detection and Mitigation: Monitor real-time HPC data, predict behavior, calculate RED, and use the KS test for anomaly detection. If an anomaly is detected, an alarm is triggered, and a mitigation strategy (like switching to a safe, albeit potentially older, version of the control code) can be initiated.

Datasets

The paper does not explicitly mention publicly available datasets for training or evaluation. Instead, it describes experiments conducted on a real power-grid controller (a Wago PLC with an ARM Cortex A8 processor running a custom real-time Linux OS). The "datasets" used were generated by:

  • Collecting normal HPC readings from legitimate processes running on the controller in a clean environment.
  • Simulating six representative real attacks against the PLC, including variations of known attacks like Stuxnet and BlackEnergy, to evaluate the detection capabilities.

Key Insights and Ideas

  • HPCs as a Stealthy Indicator: Leveraging low-level HPCs provides a robust and tamper-resistant way to monitor controller behavior, bypassing the limitations of software-based security and the risks of data exfiltration.
  • RED for Robust Anomaly Detection: Focusing on the distribution of reconstruction errors (RED) rather than individual error points makes the detection more resilient to noise and less sensitive to the exact accuracy of the prediction model. This is crucial for zero-day attacks where precise prediction of malicious behavior is impossible.
  • Zero-Day Attack Detection without Prior Knowledge: The data-driven, learning-based approach, trained solely on normal behavior, inherently allows for the detection of novel and unknown attacks.
  • Temporal Deep Learning for Dynamic Behavior: LSTM networks are well-suited for capturing the sequential and temporal nature of controller operations, enabling a more accurate model of normal behavior.
  • KS Test for Distributional Shift: The non-parametric KS test is a powerful tool for comparing probability distributions without making assumptions about their underlying form, making it ideal for comparing REDs.

Key Results

The proposed LSTM+RED system demonstrated exceptional performance:

  • High Accuracy: Achieved an accuracy of 99.97%.
  • Near-Zero False Positives: Reported a false positive rate of 0.0005%.
  • No False Negatives: All evaluated attacks were successfully detected.
  • Low Latency: The total detection latency was <360ms, which is significantly faster than human response times, allowing for quick mitigation.
  • Superiority over Conventional Methods: Outperformed various conventional anomaly detection algorithms (KNN, PCA, Isolation Forest, etc.) in terms of accuracy and F1 score, particularly because these methods often require handcrafted features that are difficult to derive from HPC data.
  • Effectiveness of RED: The RED enhancement significantly improved the performance of different LSTM architectures, making the detection more robust and less dependent on the specific deep learning model used.

Limitations and Future Work

  • HPC Availability: The system relies on the availability of HPCs. If the PLC is shut down, HPC readings are unavailable, or the HPC reading process is terminated by a malicious process, detection will fail.
  • Model Evolution: To accommodate legitimate changes in controller behavior (e.g., code updates), the deep learning model needs to be retrained or fine-tuned online. This retraining must be done using data from a clean environment.
  • Distinguishing Attacks from Benign Anomalies: The paper notes that the system detects "anomalous behavior," which could be an attack or a benign but unusual event. Further situation-dependent testing would be needed to definitively classify detected anomalies.
  • Future Work: The authors suggest further investigation into distinguishing between attacks and benign anomalies, and exploring more advanced deep learning architectures. They also highlight the need for robust mitigation strategies.

Summary updated

Abstract

Controllers of security-critical cyber-physical systems, like the power grid, are a very important class of computer systems. Attacks against the control code of a power-grid system, especially zero-day attacks, can be catastrophic. Earlier detection of the anomalies can prevent further damage. However, detecting zero-day attacks is extremely challenging because they have no known code and have unknown behavior. Furthermore, if data collected from the controller is transferred to a server through networks for analysis and detection of anomalous behavior, this creates a very large attack surface and also delays detection. In order to address this problem, we propose Reconstruction Error Distribution (RED) of Hardware Performance Counters (HPCs), and a data-driven defense system based on it. Specifically, we first train a temporal deep learning model, using only normal HPC readings from legitimate processes that run daily in these power-grid systems, to model the normal behavior of the power-grid controller. Then, we run this model using real-time data from commonly available HPCs. We use the proposed RED to enhance the temporal deep learning detection of anomalous behavior, by estimating distribution deviations from the normal behavior with an effective statistical test. Experimental results on a real power-grid controller show that we can detect anomalous behavior with high accuracy (>99.9%), nearly zero false positives and short (<; 360ms) latency.
BibTeX
@inproceedings{he2019power,
  author = {Zecheng He and Aswin Raghavan and Guangyuan Hu and Sek M. Chai and Ruby B. Lee},
  title = {{Power-Grid Controller Anomaly Detection with Enhanced Temporal Deep Learning}},
  booktitle = {TrustCom/BigDataSE},
  year = {2019},
  url = {https://openreview.net/forum?id=Y0DpknuY8J},
}
Figure