Quantization and Pruning Under FHE Constraints: Enabling Encrypted Anomaly Detection in Smart City IoT

Anomaly detection in urban IoT networks is essential for the resilience and security of smart cities. However, the privacy of sensor data remains a critical concern. In this paper, we present a machine learning-based anomaly detection approach that operates entirely on homomorphically encrypted data using Concrete ML. To enable practical deployment under the computational constraints of Fully Homomorphic Encryption (FHE), we investigate the impact of quantization-aware training (QAT) and unstructured pruning, two established techniques for reducing model complexity. We evaluate different combinations of quantization levels and pruning degrees with respect to classification performance and resource efficiency. Our results indicate that competitive accuracy can be retained even under aggressive model compression, enabling efficient and privacypreserving inference.