基于多通道融合卷积的联邦学习入侵检测模型

Federated Learning Intrusion Detection Model Based on Multi-channel Fusion Convolution

随着新一代信息技术向各行业融合渗透,经济社会数字化转型迫在眉睫。近年来,随着农业4.0时代的到来,农业领域也面临着新的机遇和挑战,农业物联网的大规模部署在带来诸多便利的同时,网络安全问题也随之出现。传统入侵检测设备无法有效应对各类未知攻击,而数据孤岛现象及数据隐私保护需求更是加剧了这一挑战,为此,文中创新性地提出了一种基于多通道融合卷积的联邦学习农业物联网入侵检测模型,旨在解决数据孤岛与数据隐私保护之间的矛盾,同时提高模型的准确率。模型的数据处理模块采用生成对抗网络对欠采样数据进行扩充,数据分析模块采用横向联邦学习机制,服务端选用联邦平均算法,客户端采用一维多通道融合卷积网络,利用多个不同尺寸的卷积核对同一段数据进行处理,捕捉不同尺度下的特征信息,再将这些特征进行融合,有效保留流量关键特征。实验结果表明,该模型在CIC-IDS2017数据集上可以实现98%的精度,并在初始阶段快速收敛,经过10轮训练后,其趋于稳定,达到99.72%的准确率和F1分数。

With the integration of next-generation information technologies across various industries,the digital transformation of economic and social sectors has become imperative.In recent years,with the advent of the Agriculture 4.0 era,the agricultural sector faces new opportunities and challenges.While the large-scale deployment of the Agricultural Internet of Things brings many conveniences,network security issues have also emerged.Traditional intrusion detection devices are inadequate in effectively handling various unknown attacks.The phenomena of data silos and the need for data privacy protection exacerbate these challenges.To address this,the paper innovatively proposes a federated learning intrusion detection model for agricultural Io T based on multi-channel fusion convolution.The model aims to resolve the contradictions between data isolation and data privacy protection while enhancing the model's accuracy.The data processing module of the model utilizes generative adversarial networks to augment undersampled data.The data analysis module adopts a horizontal federated learning mechanism.The server-side employs a federated averaging algorithm,and the client-side utilizes a one-dimensional multi-channel fusion convolutional network.This setup processes the same data segment with convolutional kernels of various sizes to capture feature information at different scales.These features are then fused together,effectively preserving key traffic characteristics.The experimental results show that the model can achieve an accuracy of 98% and converge quickly in the initial stage on the CIC-IDS2017 dataset.After 10 rounds of training,it tends to stabilize,achieving an accuracy of 99.72% and F1 score.