Scholarlink

The complexity of modern networks necessitates advanced intrusion detection, as traditional rule-based approaches struggle against evolving attacks. Autoencoders offer a promising unsupervised solution by modeling normal behavior and detecting anomalies. However, most studies use predefined architectures, neglecting optimization’s impact on performance. This study introduces a memetic algorithm-based approach that combines a genetic algorithm for global search with simulated annealing for local refinement, optimizing the autoencoder’s structure. The public Zeek datasets were used to evaluate the performance of the proposed Model. The results reveal that the proposed model using a 99th percentile threshold, achieved zero false negatives and minimizes false positives, making it a practical intrusion detection system (IDS). Future work includes adaptive thresholding and federated learning for distributed IDS.

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

Optimizing Autoencoder Architectures for Network Anomaly Detection Using Memetic Algorithm

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Omar Alhammadi,
  • Sohail Abbas,
  • Ayad Turky

摘要

The complexity of modern networks necessitates advanced intrusion detection, as traditional rule-based approaches struggle against evolving attacks. Autoencoders offer a promising unsupervised solution by modeling normal behavior and detecting anomalies. However, most studies use predefined architectures, neglecting optimization’s impact on performance. This study introduces a memetic algorithm-based approach that combines a genetic algorithm for global search with simulated annealing for local refinement, optimizing the autoencoder’s structure. The public Zeek datasets were used to evaluate the performance of the proposed Model. The results reveal that the proposed model using a 99th percentile threshold, achieved zero false negatives and minimizes false positives, making it a practical intrusion detection system (IDS). Future work includes adaptive thresholding and federated learning for distributed IDS.