Federated Learning (FL) has emerged as a revolutionary paradigm for privacy-preserving machine learning across distributed Internet of Things (IoT) environments. However, contemporary federated learningbased intrusion detection systems encounter challenges such as non-IID data distributions, class imbalance, limited communication resources, and the progression of cyber threats. This paper introduces a novel hybrid federated learning framework that integrates three critical components: TabTransformer for optimal tabular data representation, FedMADE Weighting for adaptive client aggregation employing Class Probability Matrices (CPM) and client clustering through DBSCAN via a systematic optimization process, and Gradient Similarity Aggregation (GSA) for effective outlier filtration. The suggested technique is evaluated using the comprehensive CICIoT2023 dataset, comprising over 46 million samples and 34 unique attack types. Experimental results demonstrate outstanding performance, achieving 99.41% accuracy in assessments. This research introduces a scalable, efficient, and resilient federated learning approach tailored for actual IoT intrusion detection applications.