Topology-Aware Quantization and Pruning for FHE-Based Anomaly Detection in Smart City IoT

Security and resilience are paramount in urban IoT networks, yet stringent privacy requirements for sensor data remain a significant challenge to large-scale deployment. In this extended study, we present a comprehensive privacy-preserving anomaly detection framework that operates entirely on homomorphically encrypted data. To bridge the gap between the theoretical promise of Fully Homomorphic Encryption (FHE) and the practical constraints of deployment, we investigate model compression techniques, specifically Quantization-Aware Training (QAT) combined with unstructured and structured pruning, as a means to navigate the complex design space created by the intersection of IoT limitations, FHE constraints and machine learning (ML). We systematically evaluate a wide range of compression configurations across two realistic IoT intrusion detection datasets and quantify their joint impact on classification performance, encrypted inference latency, and deployment footprint. Specifically, we identify a critical quantization threshold: 3-bit models achieve encrypted inference in 1–18 s per sample, while the transition to 5-bit precision triggers a  ∼ 100 ×  latency increase with  ≤ 3 percentage points gain in weighted F1. Topology-aware structured pruning further reduces latency by 4–6 ×  relative to parameter-matched unstructured counterparts. These results delineate a practical design corridor for FHE-enabled IoT anomaly detection and demonstrate that principled compression can bring privacy-preserving inference within operationally acceptable bounds.