Research On Collaborative Intrusion Detection Technology Based On Federated Learning

With the development of network technology,the problem of data islands is seriously hindering the growth of the artificial intelligence industry.As an application of artificial intelligence in network security,collaborative intrusion detection also faces the problem of how to safely aggregate network traffic data from multiple parties.Federated learning has a wide range of applications in privacy computing and security aggregation.It can effectively solve the problem of data islands,but it also has issues such as high-security costs,hidden security risks in the framework itself,and poor robustness.Therefore,reducing the cost of federated learning security,improving the adaptability of the training model,and improving the reliability and performance of collaborative intrusion detection have essential research value.The thesis takes joint intrusion detection as the breakthrough point and conducts in-depth research on the existing security and performance problems.First of all,to address the problem of federated GAN attacks,based on the idea of counterattack combined with loose differential privacy protection technology,a GAN-based data protection scheme is proposed.They blocked the attacker during the training process,eliminated traditional defense thinking based on data encryption,and effectively reduced the security cost.In the simulated attack and defense experiments,the experimental results show that the data protection scheme can complete the data protection simultaneously as the standalone training and improve the detection effect of rare samples.Secondly,in response to the performance problem of the proposed data protection scheme,an intermittent training method based on asynchronous training is proposed,which improves the training speed through asynchronous means and minimizes the impact of the data protection scheme.In the performance test ablation experiment,the asynchronous optimization scheme dramatically reduces the training time under the premise of ensuring the training effect,which is almost the same as the training time of the stand-alone machine.Finally,to solve the problem of slow convergence of federated learning under particular data distribution or poor model applicability,combined with data volume optimization based on FedAvg,weights are assigned to different participants through data normalization,and each participant is balanced.In combination with the pre-training scheme,the model converges quickly.Under particular data distribution,the convergence speed is greatly improved compared with FedAvg.The scheme proposed in this paper can solve the problem of collaborative intrusion detection,improve the training effect from both security cost and training efficiency.Apply to the scenario of multi-same equipment collaborative modeling,and realize the efficiency of collaborative modelling from many angles,which has availability and efficiency.