Ai-enabled framework for anomaly detection in power distribution networks under false data injection attacks
Hasnain Ahmad, Ghulam Mustafa, Muhammad Majid Gulzar, Ijaz Ahmed, Muhammad Khalid
Abstract
Modern power distribution networks are becoming cyber physical systems due to the addition of more advanced metering infrastructure (AMI). This has introduced new vulnerabilities to cyber threats, particularly false data injection (FDI) attacks. These attacks compromise the integrity of power consumption data, leading to financial losses, operational inefficiencies, and grid instability. Rule-based techniques and traditional machine learning models are two examples of traditional anomaly detection methods that often have problems. Often, these methods generate an excessive number of false alarms, struggle to adapt to new attack patterns, and perform poorly in large-scale deployments. This research suggests a robust anomaly identification framework (AIF) that uses an autoencoder (AE) for feature transformation and a multi-layer perceptron (MLP) to identify anomalies in AMI integrated with smart grids. The proposed approach first applies synthetic features extraction inspired by real-world smart meter capabilities and transforms the dataset using a denoising AE. MLP assisted in the classification to detect multiple FDI attack types with improved accuracy and reliability. Numerous experiments have been performed, and the results indicate that the suggested method works better than popular methods like correlation analysis, techniques based on clustering, and standard outlier identification algorithms. Compared to baseline methods, the proposed technique improves detection accuracy by up to approximately 25%, reduces false positives, and enhances the system’s ability to generalize across different cyberattack strategies. The proposed work computes seven different types of criterion matrices to verify the effectiveness of finding anomalies. The overall average results include mean squared error (0.0793), accuracy (92%), F1-Score (92%), recall (91%), specificity (94%), area under the curve (97%), and mean average precision (96%). These findings accentuate the potential of the proposed AIF performance in fortifying smart grid cybersecurity.