Performance Guaranteed Deep Learning for Detection of Cyber-Attacks in Dynamic Smart Grids

False Data Injection Attacks (FDIAs) pose a critical threat to the reliability of power systems, especially under dynamic operational conditions like line outages that cause data distribution changes and concept drift. Traditional supervised methods depend on labeled datasets, which are costly and impractical for real-time application, and often fail to adapt to new attack vectors and operational changes without extensive retraining. To address these challenges, we design the Deep Contrastive Variational Network (DCVN), an unsupervised learning framework engineered to detect FDIA without requiring labeled data or assumptions about network topology. The DCVN framework starts with a Deep Belief Network (DBN) for robust feature extraction from raw power system data, capturing underlying patterns in an unsupervised manner. Recognizing the limitations of DBNs in managing FDIA complexities under varying conditions, we enhance our model by integrating a modified Variational Autoencoder (VAE) with Noise Contrastive Estimation (NCE). This integration creates a novel loss function that optimizes the latent space of the VAE to enhance the contrast between normal operations and anomalies. The NCE component specifically sharpens the model's sensitivity to anomalies by maximizing the distinguishability of data representations. This design enables the DCVN to dynamically learn robust features, not just by minimizing reconstruction error but by enhancing anomaly detection through unsupervised learning. We provide a theoretical foundation for our method, ensuring performance guarantees in the dynamic environments of power grids. Empirical results demonstrate that the DCVN model significantly surpasses traditional approaches, offering robust detection capabilities.