Subramanian G, Chinnadurai M
Department of Computer Science and Engineering, E.G.S. Pillay Engineering College, Nagapattinam, Tamil Nadu, 611002, India.
Sci Rep. 2024 Dec 30;14(1):32038. doi: 10.1038/s41598-024-83682-z.
Cyber-attack brings significant threat and become a critical issue in the digital world network security. The conventional procedures developed to detects are centralized and often struggles with concerns like data privacy and communication overheads. Due to this, conventional methods are unable to adapt quickly for different threats. This research aims to develop a novel solution to address these limitations through Federated Learning. The centralized approach is developed by integrating spatio-temporal attention network and also introduces a quantum inspired federated averaging optimization procedure for cyber-attack detection. The presented model utilizes a hierarchical model aggregation procedure which dynamically groups nodes into regions based on the network condition and data similarity. A robust global model is generated at the central server by aggregating intermediate models which are developed using weighted local models. Additionally, a multi-stage model refinement procedure and privacy preservation techniques are incorporated to improve overall security and performance. The novel STAN used in the proposed work captures the spatio-temporal patterns in the network traffic data. The optimization model QIFA utilizes quantum principles to enhance the federated learning procedure. Experimentation of the proposed model utilizes benchmark UNSW-NB15 dataset and evaluated the proposed model performances. The proposed model attained better performance in detecting different types of anomalies. With maximum precision of 98.2%, recall of 98.5%, f1-score of 98.35%, specificity of 98.2% and accuracy of 98.34%, the proposed model performs better than traditional CNN, LSTM, RNN and federated learning models.
网络攻击带来了重大威胁，成为数字世界网络安全中的一个关键问题。为检测而开发的传统程序是集中式的，并且经常在数据隐私和通信开销等问题上挣扎。因此，传统方法无法快速适应不同的威胁。本研究旨在通过联邦学习开发一种新颖的解决方案来解决这些局限性。通过集成时空注意力网络开发了集中式方法，并引入了一种受量子启发的联邦平均优化程序用于网络攻击检测。所提出的模型利用分层模型聚合程序，该程序根据网络条件和数据相似性将节点动态分组到区域中。通过聚合使用加权局部模型开发的中间模型，在中央服务器上生成一个强大的全局模型。此外，还纳入了多阶段模型细化程序和隐私保护技术，以提高整体安全性和性能。在所提出的工作中使用的新颖的时空注意力网络（STAN）捕捉网络流量数据中的时空模式。优化模型量子启发联邦平均（QIFA）利用量子原理来增强联邦学习程序。对所提出模型的实验使用了基准UNSW-NB15数据集，并评估了所提出模型的性能。所提出的模型在检测不同类型的异常方面取得了更好的性能。所提出的模型的最大精度为98.2%，召回率为98.5%，F1分数为98.35%，特异性为98.2%，准确率为98.34%，其性能优于传统的卷积神经网络（CNN）、长短期记忆网络（LSTM）、循环神经网络（RNN）和联邦学习模型。
