Scholarlink

With the rapid growth and adoption of mobile networks. 5G networks have become the most popular amongst users. But, that also means the attack surface has increased for cyber threats, which is why it is necessary to implement more adaptive and intelligent Intrusion Detection Systems. Traditional IDS and deep learning models often require full retraining when new data is introduced, leading to catastrophic forgetting where previously learned knowledge is forgotten, and the model must start from scratch. This research investigates three continual learning strategies that will help mitigate catastrophic forgetting in neural networks first being Learning without Forgetting (LwF), Elastic Weight Consolidation (EWC), and Experience Replay (ER), which will then be used on the 5G-NIDD dataset. This experiment will be measured in quantitative evaluation using metrics such as Accuracy, F1-score, AUC, and forgetting rate. The results concluded that Experience Relay had consistently outperformed other models, displaying the highest F1 score and lowest forgetting rate. This means that Experience Relay provides a more resilient approach for incremental learning in IDS by ensuring both new and previously learned attacks are being detected. This research will contribute to the development of robust cybersecurity solutions for the next-generation network technology and provide a benchmark for future studies in continual learning for 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

Reducing Catastrophic Forgetting in Incremental Learning Models for 5G Malicious Traffic Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Roska Takayawa,
  • Tushaar Sharma,
  • Mansour H. Assaf,
  • Bibhya Sharma

摘要

With the rapid growth and adoption of mobile networks. 5G networks have become the most popular amongst users. But, that also means the attack surface has increased for cyber threats, which is why it is necessary to implement more adaptive and intelligent Intrusion Detection Systems. Traditional IDS and deep learning models often require full retraining when new data is introduced, leading to catastrophic forgetting where previously learned knowledge is forgotten, and the model must start from scratch. This research investigates three continual learning strategies that will help mitigate catastrophic forgetting in neural networks first being Learning without Forgetting (LwF), Elastic Weight Consolidation (EWC), and Experience Replay (ER), which will then be used on the 5G-NIDD dataset. This experiment will be measured in quantitative evaluation using metrics such as Accuracy, F1-score, AUC, and forgetting rate. The results concluded that Experience Relay had consistently outperformed other models, displaying the highest F1 score and lowest forgetting rate. This means that Experience Relay provides a more resilient approach for incremental learning in IDS by ensuring both new and previously learned attacks are being detected. This research will contribute to the development of robust cybersecurity solutions for the next-generation network technology and provide a benchmark for future studies in continual learning for IDS.