Supervisor:
Southwest Ordnance Industry Bureau
Organizer:
Chongqing Ordnance Industry Society
Chongqing University of Technology
Chongqing University of Technology
Research on network anomaly detection based on SOINN incremental autoencoders
| DOI: | 10.11809/bqzbgcxb2023.05.034 |
| Keywords: | anomaly detection; self organizing incremental neural network; automatic encoder; incremental learning; online learning |
| Abstract: | Aiming at the problem that the network anomaly detection model of batch learning has large memory resource consumption and cannot be updated online, this paper proposes an incremental autoencoder construction method by using the incremental learning characteristics of self organizing incremental neural network (SOINN). The output neuron of the improved SOINN is used as the input of the autoencoder so that the model can be updated incrementally without destroying the existing learning results. Concerning that the learning rate of the neighbor node of the winning neuron in SOINN algorithm is fixed, which is not conducive to distinguishing its similarity with the input sample, an adaptive adjustment method of the learning rate is proposed to improve the learning efficiency of the neighbor node of the winning neuron so that the output neurons of the algorithm can better represent the sample characteristics. As the purity of normal samples in the feedback update samples are not high, a sample label screening mechanism based on distance measurement is proposed, which filters normal samples by calculating the distance between feedback samples and neurons. In this way, the proportion of normal samples in the feedback samples is higher so as to improve the online detection effect of the model. The experiments are carried out on the NSL KDD dataset, the results of which prove that the proposed method has incremental learning ability. In addition, the incremental learning effect of the improved SOINN is better than that of the original algorithm, which effectively saves the computational and storage costs of the model. The online detection ability of the model improves effectively through the distance based sample label screening mechanism. |
| Issue: | Vol. 44 No. 5 (2023) |
| Published: | 2023-05-28 |
| PDF HTML | |