Intrusion Detection Systems (IDS) are vital for monitoring and securing computer networks amidst the increasing number of cyber-attacks. Intelligent IDS systems utilizing Machine Learning (ML) and Deep Learning (DL) offer high classification accuracy but traditionally require centralized data, leading to concerns about privacy breaches and data transmission costs. Federated Learning (FL) addresses these issues through a privacy-preserving, decentralized approach that keeps data local while sharing only model parameters with a central server. This study investigates an FL-based IDS for Internet of Medical Things (IoMT), addressing challenges like data heterogeneity and malicious client poisoning attacks. Authors propose a personalized FL-based IDS (pFL-IDS) to manage imbalanced data distributions in heterogeneous IoMT datasets. The study demonstrates that malicious clients can significantly degrade the performance of an FL-based IDS. To counteract this, pFL-IDS employs a Bayesian Logit Calibration (BLC) for local model training and a server-side detection mechanism to identify malicious clients. This mechanism compares local models with a non-poisoned centroid determined by the pre-computed global model. pFL-IDS effectively detects poisoning attacks and outperforms baseline methods without sacrificing performance.