In smart power grids, electricity theft causes huge economic losses to electrical utility companies. Machine learning (ML), especially deep neural network (DNN) models hold state-of-the-art performance in detecting electricity theft cyberattacks. However, DNN models are vulnerable to adversarial attacks, i.e., evasion attacks. In this work, we, first, study the vulnerability of the DNN-based electricity theft detectors against evasion attacks and the influence of the model’s regularization (generalization) and transferability on robustness. Then, we cluster the power consumers and train a detector for each cluster, and compare the performance and robustness of this detector to a global detector that is trained on all the consumers’ data. The results indicate that the cluster-based detector is not only more robust against evasion attacks but also enhances normal classification accuracy because its training data has more consumption pattern similarity compared to the training data of the global detector which requires higher level of regularization. Then, we develop a cluster-based parallel-ensemble electricity theft detector to achieve high robustness against evasion attacks and high detection accuracy. The proposed detector has two levels of defense, including clustering and ensemble, where clustering decreases regularization and improves robustness, and the ensemble of diverse decision models improves robustness against transferability. We evaluate the detector on different threat models, including Blackbox and Graybox with different knowledge about defense strategy, model architecture, and training dataset. The evaluation results indicate that unlike the existing solutions that sacrifice the normal accuracy of the model to improve the robustness against evasion attacks, the proposed detector outperforms the benchmark defenses in terms of robustness and normal classification accuracy.

In the realm of smart grids, machine learning (ML) detectors—both binary (or supervised) and anomaly (or unsupervised)—have proven effective in detecting electricity theft (ET). However, binary detectors are designed for specific attacks, making their performance unpredictable against new attacks. Anomaly detectors, conversely, are trained on benign data and identify deviations from benign patterns as anomalies, but their performance is highly sensitive to the selected threshold values. Additionally, ML detectors are vulnerable to evasion attacks, where attackers make minimal changes to malicious samples to evade detection. To address these limitations, we introduce a hybrid anomaly detector that combines a Deep Auto-Encoder (DAE) with a One-Class Support Vector Machine (OCSVM). This detector not only enhances classification performance but also mitigates the threshold sensitivity of the DAE. Furthermore, we evaluate the vulnerability of this detector to benchmark evasion attacks. Lastly, we propose an accurate and robust cluster-based DAE+OCSVM ET anomaly detector, trained using Explainable Artificial Intelligence (XAI) explanations generated by the Shapley Additive Explanations (SHAP) method on consumption readings. Our experimental results demonstrate that the proposed XAI-based detector achieves superior classification performance and exhibits enhanced robustness against various evasion attacks, including gradient-based and optimization-based methods, under a black-box threat model.

In the realm of smart grids, smart meters can be hacked to report false data to lower the consumers’ electricity bills. While machine learning (ML) techniques have shown promise in detecting false data, they are also prone to adversarial attacks, such as evasion attacks. This article investigates the impact of gradient-ensemble-based evasion attacks on the smart grid ML-based false data detectors, focusing on the white-box threat model where attackers possess detailed knowledge of the defense mechanism. First, we examines the vulnerability of three detectors (consumer-based, cluster-based, and global) to gradient-based evasion attacks. The evaluation results show an inverse relationship between robustness of the detectors and regularization (i.e., generalization), where higher data set variability usually causes higher regularization. Notably, minimal regularization level is observed when electricity consumption patterns are close. Our findings also indicate that the consumer-based detector exhibits higher accuracy and robustness but remains susceptible to zero day attacks and demands substantial computational resources for training an ML model for each consumer. In contrast, the cluster-based detector improves accuracy and exhibits satisfactory robustness compared to the global detector. Subsequently, we proposes two parallel-ensemble approaches (stacking and voting) for the cluster-based false data detectors trained on the adversarial samples. The evaluation results demonstrate that integrating clustering, adversarial training, and ensemble methods, the proposed detector enhances robustness against gradient-ensemble-based evasion attacks while significantly boosting accuracy. This stands in contrast to benchmark defenses, which often face a tradeoff between accuracy and robustness, sacrificing accuracy to bolster resilience against evasion attacks.

Reinforcement learning (RL) is proven effective in optimizing home battery charging coordination within smart grids. However, its vulnerability to adversarial behavior poses a significant challenge to the security and fairness of the charging process. In this study, we, first, craft five stealthy false data injection (FDI) attacks that under-report the state-of-charge (SoC) values to deceive the RL agent into prioritizing their charging requests, and then, we investigate the impact of these attacks on the charging coordination system. Our evaluations demonstrate that attackers can increase their chances of charging compared to honest consumers. As a result, honest consumers experience reduced charging levels for their batteries, leading to a degradation in the system’s performance in terms of fairness, consumer satisfaction, and overall reward. These negative effects become more severe as the amount of power allocated for charging decreases and as the number of attackers in the system increases. Since the total available power for charging is limited, some honest consumers with genuinely low SoC values are not selected, creating a significant disparity in battery charging levels between honest and malicious consumers. To counter this serious threat, we develop a deep learning-based FDI attack detector and evaluated it using a real-world dataset. Our experiments show that our detector can identify malicious consumers with high accuracy and low false alarm rates, effectively protecting the RL-based charging coordination system from FDI attacks and mitigating the negative impacts of these attacks.

A major challenge in smart power grids (SGs) is the injection of false data by malicious actors, which may disrupt the proper operation and cause financial losses. To detect such false data, machine learning (ML)-based methods are commonly used, with models being trained through either centralized learning (CL) or federated learning (FL) approaches. While CL approach achieves higher classification accuracy, it introduces significant privacy risks. In contrast, the existing FL approaches mitigate privacy concerns by keeping data decentralized but suffer from lower classification accuracy compared to CT approach in addition to high overhead in terms of communication. To address these limitations, we propose an explainable artificial intelligence (XAI)-based FL approach that is trained on the explanations of electricity power consumption readings instead of the readings themselves. This approach not only improves false data detection accuracy but also significantly reduces FL communication overhead. The idea is that by leveraging XAI, our approach can effectively distinguish between benign and malicious samples, enabling the establishment of a precise decision boundary with fewer training rounds. To verify our approach, we conducted extensive evaluations using two real-world -electricity consumption datasets under various cyber-attack scenarios. The experimental results demonstrate that our XAI-based FL approach outperforms the existing data-based FL approaches, achieving superior accuracy and reduced overhead. For instance, in some cases, our approach was able to reduce the number of FL rounds by more than 80%, reduce the false alarm by more than 82%, and increase the detection accuracy by more than 14%.

In smart grids, smart meters (SMs) transmit power consumption data to utilities for billing and energy management. However, compromised SMs can report low consumption to reduce electricity bills. Deep reinforcement learning (DRL) detectors have recently been proposed to detect these attacks due to their adaptability to new attacks and changes in power consumption patterns. This article explores backdoor attacks targeting DRL detectors during training, aiming to introduce a vulnerability in the detector. These attacks make the detector misclassify false low-consumption data when trigger samples are used while maintaining normal classification accuracy otherwise. We propose a DRL-based attack model that generates stealthy and unique trigger samples using cosine similarity. Our evaluations show the attack is initially highly successful, but its success diminishes with honest data used for incremental training of the detector. To sustain high success rates, attackers must influence incremental training. We also propose defenses, including data filtration during the preparation stage, adversarial training for the defense model during the training stage, and a combined approach, with experiments validating their effectiveness.