In the advanced metering infrastructure (AMI) of the smart grid, smart meters (SMs) are deployed to collect fine-grained electricity consumption data, enabling billing, load monitoring, and efficient energy management. However, some consumers engage in fraudulent behavior by hacking their meters, leading to either traditional electricity theft or more sophisticated evasion attacks (EAs). EAs aim to illegally reduce electricity bills while deceiving theft detection mechanisms. The current methods for identifying such attacks raise privacy concerns due to the need for access to consumers’ detailed consumption data to train detection mechanisms. To address privacy concerns, federated learning (FL) is proposed as a collaborative training approach across multiple consumers. Adversarial training (AT) has shown promise in countering evasion threats on machine learning models. This paper, first, investigates the susceptibility of traditional electricity theft classifiers trained by FL to EAs for both independent and identically distributed (IID) and Non-IID consumption data. Then, it investigates the effectiveness of AT in securing the global electricity theft detector against EAs, assuming no misbehavior from the participant consumers in the FL process. After that, we introduce three novel attacks, namely Distillation , No-Adversarial-Sample-Training , and False-Labeling , which can be launched during the AT process to make the global model susceptible to evasion at inference time. Finally, extensive experiments are conducted to validate the severity of these proposed attacks. Our findings reveal that the AT can counter EAs effectively when the FL participants are honest, but it fails when they act maliciously and launch our attacks. This works lays the foundation for future endeavors in exploring additional countermeasures, in conjunction with AT, to bolster the security and resilience of FL machine learning models against adversarial attacks in the context of electricity theft detection.

Existing literature confirms the ability of machine learning to identify fraudulent smart grid power consumers who report false consumption readings to pay less electricity bills. Additionally, federated learning (FL) shows promise as a way to train the detection model without requiring data sharing, thereby safeguarding consumer privacy. However, malicious participants (i.e., clients) in FL training can launch adversarial attacks by training their local models with specially crafted low-consumption data to inject a Trojan into the global model. This Trojan can then be activated during the evaluation phase to evade the detection of false data. To the best of our knowledge, not enough research has been done on this topic in the context of unsupervised learning. The absence of labels in unsupervised learning exacerbates the effectiveness of Trojan attacks and renders it more challenging to design robust defense mechanisms. In this article, we first investigate the vulnerability of one-class classifiers to Trojan attacks. Then, we propose two defense approaches named layerwise close-to-median (LWCM) and Machine Unlearning to counter this attack. In LWCM, by choosing a FL client whose last layer model parameters are near to the median of all clients’ last layer parameters to update the global model, we can identify and exclude malicious updates. The idea is that the last layer parameters of honest clients should be similar, whereas those from malicious clients are different. With the majority of clients being honest, the median values are closer to the parameters of these clients, facilitating the detection of malicious clients. In Machine Unlearning, we utilize gradient ascent-based techniques to adapt models by selectively removing attacker-related data points. This is possible because honest clients generate data resembling that of malicious clients and employ a dual-component loss function to maintain model proficiency in recognizing benign power consumption patterns while eliminating malicious patterns. To show the seriousness of Trojan attacks and the effectiveness of our countermeasures, many experiments have been carried out.

In smart power grid, consumers can hack their smart meters to report low electricity consumption readings to reduce their bills launching electricity theft cyberattacks. This study investigates a Trojan attack in federated learning of a detector for electricity theft. In this attack, dishonest consumers train the detector on false data to later bypass detection, without degrading the detector’s overall performance. We propose three defense strategies: Redundancy, Med-Selection and Combined-Selection . In the Redundancy approach, redundant consumers with similar consumption patterns are included in the federated learning process, so their correct data offsets the attackers’ false data when the local models are aggregated. Med-Selection selects the median model parameters of consumers with similar usage patterns to reduce outlier influence. In Combined-Selection , we compare gradients from consumers with same consumption patterns to the median of all local models, leveraging the fact that honest consumers’ gradients are closer to the median while malicious ones deviate. Our experiments using real-world data show the Trojan attack’s success rate can reach 90%. However, our defense methods reduce the attack success rate to about 7%, 4%, and 3.3% for Redundancy, Med-Selection , and Combined-Selection , respectively, when 10% of consumers are malicious.

The advanced metering infrastructure (AMI) of the smart grid plays a critical role in energy management and billing by enabling the periodic transmission of consumers’ power consumption readings. To optimize data collection efficiency, AMI employs a “change and transmit” (CAT) approach. This approach ensures that readings are only transmitted when there is enough change in consumption, thereby reducing data traffic. Despite the benefits of this approach, it faces security challenges where malicious consumers can manipulate their readings to launch cyberattacks for electricity theft, allowing them to illegally reduce their bills. While this challenge has been addressed for supervised learning CAT settings, it remains insufficiently addressed in unsupervised learning settings. Moreover, due to the distortion introduced in the power consumption readings due to using the CAT approach, the accurate prediction of future consumption for energy management is a challenge. In this paper, we propose a two-stage approach to predict future readings and detect electricity theft in the smart grid while optimizing data collection using the CAT approach. For the first stage, we developed a predictor that is trained exclusively on benign CAT power consumption readings, and the output of the predictor is the actual readings. To enhance the prediction accuracy, we propose a cluster-based predictor that groups consumers into clusters with similar consumption patterns, and a dedicated predictor is trained for each cluster. For the second stage, we trained an autoencoder and a one-class support vector machine (SVM) on the benign reconstruction errors of the predictor to classify instances of electricity theft. We conducted comprehensive experiments to assess the effectiveness of our proposed approach. The experimental results indicate that the prediction error is very small and the accuracy of detection of the electricity theft attacks is high.

The transition from traditional power systems to smart grids demands advanced methods for detecting and classifying Power Quality Disturbances (PQDs)—variations in voltage, current, or frequency that disrupt device performance. The rise of renewable energy and nonlinear loads, such as LED lighting, has increased PQD occurrences. While deep learning models can effectively analyze data from grid sensors to detect PQD occurrences, privacy concerns often prevent operators from sharing raw data which is necessary to train the models. To address this, Federated Learning (FL) enables collaborative model training without exposing sensitive information. However, FL’s decentralized design introduces new risks, particularly data poisoning attacks, where malicious clients corrupt model updates to degrade the global model accuracy. Despite these risks, PQD classification under FL and its vulnerability to such attacks remain largely unexplored. In this work, we develop FL-based classifiers for PQD detection and compare their performance to traditionally trained, centralized models. As expected from prior FL research, we observed a slight drop in performance: the model’s accuracy decreased from 97% (centralized) to 96% (FL), while the false alarm rate increased from 0.19% to 4%. We also emulate five poisoning scenarios, including indiscriminate attacks aimed at degrading model accuracy and class-specific attacks intended to hide particular disturbance types. Our experimental results show that the attacks are very successful in reducing the accuracy of the classifier. Furthermore, we implement a detection mechanism designed to identify and isolate corrupted client updates, preventing them from influencing the global model. Experimental results reveal that our defense substantially curtails the performance degradation induced by poisoned updates, thereby preserving the robustness of the global model against adversarial influence.