The pursuit of robust and efficient federated learning systems remains pivotal as data and models become increasingly decentralised across diverse domains. This paper presents an advanced aggregation algorithm designed to enhance the security and efficiency of federated learning systems, particularly in handling malicious or misleading clients [1, 2] in independent and identically distributed (IID) and non-independent and identically distributed (non-IID) data scenarios effectively. By incorporating a game theory-inspired marginal contribution concept, which leverages a small validation set at the global model level, our algorithm aims to secure Federated Learning frameworks from attacker clients by dynamically filtering poorly-performing updates. This method selectively includes client updates, ensuring that only contributions that genuinely enhance the model’s performance are considered. As a result, the proposed algorithm shows significant improvements in aggregation accuracy and robustness in poisoned-data-and-client scenarios compared to traditional federated learning methods, as demonstrated by tests performed across various data environments.