Scholarlink

Abstract <p>Ensuring the security of large-scale computer systems is a critical issue today. This is typically achieved by analyzing user behavior based on events that occur within the system. However, current methods focus on the context of these events only and ignore the content part. Additionally, the volume of events can be significant, making it challenging to analyze all of them. This paper proposes a novel approach that utilizes parallel convolutional autoencoders and fuzzy clustering to overcome these issues. This method considers both the content and context of each event, offering a more comprehensive analysis. It provides an alternative to traditional Transformer architectures that can take a substantial amount of time to perform self-attention operations. The proposed approach demonstrates superior performance in comparison to existing techniques, making it suitable for large-scale computer security. Additionally, the paper proposes methods to optimize the efficiency of this approach by using one or more graphics processing units. This leads to a significant acceleration of training and application processes, and therefore enhances the effectiveness of the proposed solution even more.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Parallel Autoencoder for Unsupervised Anomaly Detection in Large Multimodal User Behavior Data

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • O. E. Gorokhov,
  • M. A. Kazachuk,
  • I. V. Mashechkin,
  • M. I. Petrovskiy

摘要

Abstract

Ensuring the security of large-scale computer systems is a critical issue today. This is typically achieved by analyzing user behavior based on events that occur within the system. However, current methods focus on the context of these events only and ignore the content part. Additionally, the volume of events can be significant, making it challenging to analyze all of them. This paper proposes a novel approach that utilizes parallel convolutional autoencoders and fuzzy clustering to overcome these issues. This method considers both the content and context of each event, offering a more comprehensive analysis. It provides an alternative to traditional Transformer architectures that can take a substantial amount of time to perform self-attention operations. The proposed approach demonstrates superior performance in comparison to existing techniques, making it suitable for large-scale computer security. Additionally, the paper proposes methods to optimize the efficiency of this approach by using one or more graphics processing units. This leads to a significant acceleration of training and application processes, and therefore enhances the effectiveness of the proposed solution even more.