Anti-overfitting lightweight anomaly detection neural network model retraining method

Through the lightweight anomaly detection neural network model retraining method based on deep variational autoencoder, the probability distribution of the encoder and decoder is adjusted, and the overfitting and high overhead update problems caused by data distribution changes in neural network models in cloud computing are solved, and fast and lightweight model updates and efficient anomaly detection are achieved.

WO2025043823A1PCT designated stage expired Publication Date: 2025-03-06HAINAN INST OF ZHEJIANG UNIV +1
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2023/123750
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-08-30
Filing Date
2023-10-10
Publication Date
2025-03-06

AI Technical Summary

Technical Problem

In the field of cloud computing, when neural network models are used for abnormal detection, they often overfit due to changes in data distribution, and frequent updates of models require high overhead, which cannot meet the needs of high peak traffic monitoring.

Method used

A lightweight anomaly detection neural network model retraining method based on deep variational autoencoder is proposed. By adjusting the probability distribution learned by the encoder and decoder, using the rumination module to generate shadow data, and updating the model through Bayesian formulas and mapping functions is proposed to reduce the retraining overhead and resist overfitting.

Benefits of technology

It realizes the rapid and lightweight update of the anomaly detection model when the data distribution changes, reduces the overhead of retraining, avoids overfitting, and meets the needs of high peak traffic monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2023123750_06032025_PF_FP_ABST
    Figure CN2023123750_06032025_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed in the present invention is an anti-overfitting lightweight anomaly detection neural network model retraining method. According to the method, retraining is carried out on an anomaly detection model based on a deep variational autoencoder, and when data distribution changes, conditional distribution about a hidden state and a reconstructed data sample obtained by an encoder and a decoder of the depth variational autoencoder also changes. According to the present invention, a mapping function is used to adjust conditional distribution of the hidden state and reconstructed data calculated by an old model, so as to adapt to new data distribution, the mapping function has the characteristic of being simple and convex, and when a loss function form defined by the present invention is used, the fast convergence rate and the lightweight overhead in the retraining process can be ensured. In addition, the present invention provides a rumination module for carrying out data enhancement on new observation data, and solves the problem of insufficient new observation sample data at the initial stage of changes in cloud service features.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • A training sample data expansion method and device based on a variational auto-encoder

    CN109886388A

  • Large-scale multivariate time series data anomaly detection method oriented to cloud environment

    CN112784965A

  • Depth variational auto-encoder model training method and device, equipment and storage medium

    CN113642716A

  • Method and device for updating deep learning model in network security anomaly detection

    CN116032557A

  • Data anomaly detection method and apparatus

    WO2023123941A1