Аңдатпа
Federated Learning (FL) enables distributed model training without centralizing data, but it remains vulnerable to corrupted or adversarial clients that can degrade the global model through malicious weight update. This paper proposes a modular aggregation mechanism using the Chinese Remainder Theorem (CRT) to secure model aggregation. The CRT encodes model weights under multiple modular bases, allowing consistent and tamper-resilient reconstruction of global weights. Experimental analysis demonstrates that the modular arithmetic based secure aggregation significantly mitigates the effect of corrupted clients, improving accuracy and stability compared to standard Federated Averaging (FedAvg). The modular CRTFed proposed achieves a 53.76 % reduction in the maximum validation loss. This is a great quantitative measure of improved robustness, indicating the method effectively prevents the model from diverging severely due to malicious or noisy data updates.