TADA: A Transferable Domain-Adversarial Training for Smart Grid Intrusion Detection Based on Ensemble Divergence Metrics and Spatiotemporal Features

Pengyi Liao, Jun Yan, Jean Michel Sellier, Yongxuan ZhangEnergies 2022 · DOI

TL;DR

This paper introduces TADA, a novel framework for smart grid intrusion detection that addresses the challenge of data distribution shifts. TADA uses transferability analysis to intelligently decide when to apply transfer learning and then employs a domain-adversarial approach with CNNs and LSTMs to extract robust spatiotemporal features, achieving high accuracy in detecting false data injection attacks.

What Problem Does This Paper Address?

Smart grids are critical infrastructure, but their interconnected nature makes them vulnerable to cyberattacks. Traditional machine learning models for intrusion detection often assume that training and testing data come from the same distribution. However, in real-world smart grids, this assumption rarely holds due to constantly changing system loads, topology alterations, and evolving attack strategies. These "data distribution divergences" can significantly degrade the performance of well-trained models. Furthermore, obtaining labeled attack data is difficult, making it challenging to retrain models frequently. Transfer learning (TL) offers a solution by allowing models to leverage knowledge from a source domain to a target domain. However, two key challenges remain: when is TL truly necessary to avoid performance degradation, and how can we effectively extract features that are invariant across these differing domains, especially considering the spatiotemporal nature of smart grid data?

Methods

The TADA framework tackles these challenges in two main steps:

  1. Transferability Analysis: This step aims to predict the performance drop of a trained model on a new, unseen (target) domain. The authors propose an ensemble approach that combines multiple data distribution divergence metrics (geometry-based, domain-discrimination-based, mutual-information-based, and higher-order-moment-based). These metrics are used as features to train a regression model (specifically, a neural network) that predicts the accuracy drop. This prediction helps determine if TL is needed: if the predicted drop is too small, TL is unnecessary; if it's too large, a new model might be better; and if it's in an intermediate range, TL is recommended.
  2. Spatiotemporal Domain-Adversarial Training: Once the need for TL is established, this step focuses on extracting robust features. TADA employs a domain-adversarial neural network (DANN) architecture. The core innovation here is the feature extractor, which combines Convolutional Neural Networks (CNNs) for spatial feature extraction (capturing correlations between measurements) and Long Short-Term Memory (LSTM) networks for temporal feature extraction (learning patterns over time). These spatiotemporal features are then fed into a domain discriminator, which is trained adversarially to make the features domain-invariant. This ensures that the extracted features are discriminative for attack detection but not sensitive to the differences between the source and target domains.

Datasets

The authors generated their datasets using:

  • Normal Data: Seven years of practical load demand data from ISO New England (2015-2021).
  • System Model: The standard IEEE 30-bus system was used to simulate power system measurements.
  • Attack Model: False Data Injection (FDI) attacks were synthesized by adding attack vectors to normal data, mimicking stealthy attacks that evade traditional bad data detection methods.
  • Data Generation: MATPOWER was used to generate 142 measurements per minute via DC optimal power flow analysis, with added Gaussian noise. Time series data was created using a 60-minute sliding window.
  • Domain Splits: Source domains were derived from labeled historical data (2015-2018), while target domains consisted of unlabeled data (2019-2021). Temporal variations were simulated by using different seasons (winter, spring, summer, fall) for source and target domains. Attack location variations were introduced by targeting different buses within the IEEE 30-bus system.

Key Insights and Ideas

  • Ensemble Metrics for Transferability: Relying on a single divergence metric can be insufficient. Combining multiple metrics provides a more comprehensive view of distribution differences, leading to more accurate predictions of performance degradation.
  • Intelligent TL Triggering: The framework's ability to predict when TL is beneficial is crucial. This prevents unnecessary computational overhead and potential performance issues from applying TL when it's not needed or when the divergence is too severe.
  • Spatiotemporal Invariance is Key: For smart grids, both spatial relationships between sensors and temporal sequences of measurements are vital for detecting sophisticated attacks. Combining CNNs and LSTMs within a domain-adversarial framework effectively captures and leverages these spatiotemporal characteristics while ensuring domain invariance.
  • Robustness to Imbalance: The study explicitly tested scenarios with varying attack data percentages, demonstrating that TADA maintains high performance even with imbalanced datasets, a common issue in real-world intrusion detection.

Key Results

  • Transferability Analysis Accuracy: The ensemble transferability analysis achieved a Root Mean Square Error (RMSE) of 1.79% and a Maximum Absolute Error (MaxAE) of 5.62% in predicting accuracy drops, significantly outperforming a baseline method.
  • Attack Detection Performance: The TADA framework achieved an average detection accuracy of 95.58% across various scenarios.
  • Robustness: The framework demonstrated strong robustness against variations in attack data percentages, achieving an average F1-score of 92.02%.
  • Comparison: TADA outperformed state-of-the-art TL models like DANN and CoDATS, as well as non-TL models like SVM, MLP, and FCN, in terms of both accuracy and F1-score.

Limitations and Future Work

  • Limitations: The paper acknowledges that while TADA is effective, it focuses on a specific type of TL (unsupervised TL for the target domain) and a particular attack (FDI). The robustness of the framework to other attack types or more complex domain shifts might require further investigation. The dataset generation, while realistic, is still a simulation.
  • Future Work: The authors suggest exploring more practical attack scenarios and extending the TADA framework to other Cyber-Physical Systems (CPS) beyond smart grids. Further research could also investigate more advanced TL techniques or different feature extraction methods.

Summary updated

Abstract

For attack detection in the smart grid, transfer learning is a promising solution to tackle data distribution divergence and maintain performance when facing system and attack variations. However, there are still two challenges when introducing transfer learning into intrusion detection: when to apply transfer learning and how to extract effective features during transfer learning. To address these two challenges, this paper proposes a transferability analysis and domain-adversarial training (TADA) framework. The framework first leverages various data distribution divergence metrics to predict the accuracy drop of a trained model and decides whether one should trigger transfer learning to retain performance. Then, a domain-adversarial training model with CNN and LSTM is developed to extract the spatiotemporal domain-invariant features to reduce distribution divergence and improve detection performance. The TADA framework is evaluated in extensive experiments where false data injection (FDI) attacks are injected at different times and locations. Experiments results show that the framework has high accuracy in accuracy drop prediction, with an RMSE lower than 1.79%. Compared to the state-of-the-art models, TADA demonstrates the highest detection accuracy, achieving an average accuracy of 95.58%. Moreover, the robustness of the framework is validated under different attack data percentages, with an average F1-score of 92.02%.
BibTeX
@inproceedings{liao2022tada,
  author = {Pengyi Liao and Jun Yan and Jean Michel Sellier and Yongxuan Zhang},
  title = {{TADA: A Transferable Domain-Adversarial Training for Smart Grid Intrusion Detection Based on Ensemble Divergence Metrics and Spatiotemporal Features}},
  booktitle = {Energies},
  year = {2022},
  doi = {10.3390/en15238778},
}
Figure