Mitigating Falsified Updates in Federated Learning Using a Hybrid FTWO-LSTM Framework
Abstract
Decentralized model training through Federated Learning (FL) has gained attention as an effective approach, allowing collaborative learning while avoiding the exchange of raw data. Nevertheless, this has been detected to be one of the most grave disadvantages of FL because it has been noted to be susceptible to falsified or poisoned updates by malicious clients, which drastically affects the performance, reliability, and efficiency of the global model. The current defense mechanisms are characterized by their low scalability and computational costs, which leads to a poor adversarial manipulation detection. To overcome such problems, the presented research proposes a more powerful system that combines the Fig Tree Wasp Optimization algorithm with a Long Short-Term Memory (LSTM) model to validate updates and detect anomalies in FL systems in a more robust way. The optimization element efficiently optimizes the hyperparameters which enhances the learning capability of LSTM in detecting malicious contributions. The strategy was trained and analyzed on the NSL-KDD dataset found on Kaggle, one of the established intrusion and anomaly detection benchmarks. The results of the experiments prove that the suggested framework gained an accuracy of 97.8 with the precision of 97.5, recall of 97.2, specificity of 98.1, F1-score of 97.3, and AUC-ROC of 98.4. Through comparison to the current baseline algorithms, the proposed method was found to outperform in all metrics, which points to its endeavour to enhance reliability, security, and system efficiency, in general, within federated learning settings.