Omara, Ahmed Mohamed Elsayed2024-09-262024-09-262024-09-26http://hdl.handle.net/10393/46612https://doi.org/10.20381/ruor-30575In the past decade, communication technologies such as cellular networks, Wi-Fi, and optical communication have significantly advanced, impacting daily life and enhancing urban preparedness for power outages. Smart grids, unlike traditional utility grids, enable bi-directional flows of electricity and information, improving efficiency and minimizing power losses by exchanging grid status and customer requirements. However, these advancements have also increased the attack surface, introducing new cyber vulnerabilities that adversaries can exploit, posing threats to smart grids. Our work addresses the security challenges arising from integrating Electric Vehicles (EVs), smart microgrids, and Artificial Intelligence (AI) in Vehicle-to-Microgrid (V2M) applications. First, the research investigates the growing attack surface resulting from the integration of EVs and smart grids, particularly focusing on data integrity attacks that pose a significant threat to V2M applications. A scheme leveraging unsupervised ML techniques is proposed to model and detect these attacks. Extensive simulations demonstrate the scheme’s effectiveness in reducing the impact of data integrity attacks by up to 76.5%. Next, we explore Adversarial Machine Learning (AML) attacks targeting V2M services. These attacks exploit vulnerabilities in the ML classifiers, enabling adversaries to deceive the system and disrupt microgrid operations. To anticipate and counteract these threats, we conduct an anticipatory analysis of a multi-stage attack. By simulating adversary behavior, we evaluate the robustness of the ML classifier and develop effective countermeasures. Our findings reveal that the multi-stage gray-box attack achieves an Evasion Increase Rate (EIR) of up to 73.2%, using 40% less data than traditional white-box attacks. To enhance the security of AI-based microgrid control systems in V2M services, we propose a comprehensive defense framework integrating a Generative Adversarial Network (GAN) model and a robust ML classifier. The GAN model generates realistic adversarial samples, enabling the ML classifier to learn and adapt to novel attack patterns. Additionally, the ML classifier is trained on a diverse dataset comprising both legitimate and adversarial samples, improving its ability to distinguish between normal and malicious activities. Simulations validate the effectiveness of the proposed defense mechanism, achieving an Adversarial Detection Rate (ADR) of 90.2%. To address the limited computational power and memory in V2M edge settings, we examine different model optimization techniques, such as projection, pruning, and quantization, to optimize the model’s size without compromising detection performance. The proposed method integrates model design and compression, resulting in an optimized detection model that remains robust against adversarial attacks. This approach ensures that the model remains compact and maintains high accuracy. For instance, the Convolutional Neural Network (CNN) model’s detection rate against Fast Gradient Sign Method (FGSM) attacks is 92.5% and 91% before and after compression, respectively.enAttribution-NonCommercial-ShareAlike 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-sa/4.0/Adversarial attacksevasion attackssmart meterssmart gridsedge computingcompact AIelecrtic vehiclesdata integrity attacksinference attacksvehicle-to-microgrid (V2M)vehicle-to-house (V2H)machine learningThesis