Abstract

The mushrooming of IoTs (Internet of Things) and decentralised paradigm in cyber security have attracted a lot of interest from the government, academic, and business sectors in recent years. The use of MLT-assisted techniques in the IoT security arena has attracted a lot of attention in recent years. Many current studies presume that massive training data is readily accessible from IoT devices and transferable to main servers. However, since data is hosted on single servers, security and privacy concerns regarding this data also increase. It is suggested to use decentralised on-device data in OFDL (Optimal Federated Deep Learning) based anomaly detections to proactively identify infiltration in networks for IoTs. The GRUs (Gated Recurrent Units) used in OFDL's training rounds share only learned weights with the main OFDL servers, protecting data integrity on local devices. The model's training costs are reduced by the use of appropriate parameters, which also secures the edge or IoT device. In order to optimise the hyper-parameter environments for the limited OFDL environment, this paper suggests an MSSO (Modified Salp Swarm Optimisation) approach. Additionally, ensembles combine updates from multiple techniques to enhance accuracies. The experimental findings show that this strategy secures user data privacy better than traditional/centralized MLTs and offers the best accuracy rate for attack detection.