International Journal of Science, Engineering and TechnologyAuthors: Research Scholar Sujata Pattnaik, Dr. Ridhika Sidhi
Abstract: Federated learning (FL) enables Internet of Things (IoT) clients to train a shared intrusion detection model without centralising raw network traffic, but it neither guarantees privacy nor comes free of communication cost. Model updates can leak information about local traffic, and repeated update exchange can strain constrained IoT links. This paper presents a layered privacy design for federated intrusion detection together with a communication-aware analysis of its cost. Five layers—local data retention, update clipping with differential privacy, secure aggregation, encrypted transport, and authenticated participation—are each mapped to the threat they address, the parameters that must be reported and the overhead they add. An analytic cost model combines model size, numeric encoding, participating clients and rounds. For a lightweight 5,352-parameter detector, the per-client upload per round falls from 20.9 KB with 32-bit encoding to 5.2 KB with 8-bit quantisation, compared with 195.3 KB for a 50,000-parameter convolutional model, showing that model choice and encoding dominate communication cost. The paper also specifies an ablation plan for measuring the privacy–utility–cost balance on IoT intrusion datasets.
