The increasing sophistication and scale of cyber threats necessitate intelligent and adaptive security solutions beyond traditional perimeter-based defenses. Intrusion detection systems (IDSs), particularly those leveraging machine learning (ML), have shown great promise in identifying both known and novel attack vectors. However, the centralized nature of conventional ML-based IDSs poses critical challenges, including privacy risks, communication overhead, and limited scalability across diverse edge environments. Federated learning (FL) addresses these concerns by enabling decentralized model training while preserving data privacy. Yet, FL suffers in heterogeneous environments due to non-IID data, varied client resources, and inconsistent participation. To overcome these limitations, we propose a novel FL-based IDS framework incorporating meta-sampling. This adaptive client selection mechanism leverages historical training insights to prioritize clients with representative or underrepresented data. Experimental evaluations using benchmark IDS datasets show that our framework significantly enhances convergence speed, model robustness, and detection accuracy. This study contributes toward making FL-based IDSs more effective and practical for deployment in real-world, resource-constrained cybersecurity infrastructures.