<p>The 6G network suffers from large-scale botnet attacks due to the extensive scale and complexity of the network. To address security and communication overhead issues in IoT-Edge based 6G Networks, the next-generation IDS adopts edge intelligence and federated learning architectures. In existing works, federated learning models have been implemented for botnet attacks to safeguard the privacy of IoT device data. However, these methods are based on the assumption of independent and identical data distribution (IID) of IoT devices and suffer due to non-IID and imbalanced IoT data. Another issue that has been considered in our research work is the resource constraints of IoT devices which can not perform the training of deep learning models. To solve these two issues, we present a novel approach where virtual clusters are formed at edge nodes based on the IoT data distribution. As a result, local models in the clusters converge fast towards local optima, ultimately resulting in improved overall convergence. Further, we propose a method called adaptive training epochs for localized models (ATELM) to solve the problem of imbalanced data effect on clustering. Our proposed approach enhances the classification performance while reducing the communication traffic between IoT devices and the cloud. The efficacy of the proposed approach has been shown through experiments on real-time data sets (such as N-BaIoT and Bot-IoT) and also by comparing the proposed approach with existing methods such as FedAvg and FedProx, particularly in scenarios involving non-IID and imbalanced data.</p>