Scholarlink

Distributed Federated Learning (DFL) solves the problems of single-point failure and high latency in traditional federated learning while ensuring data privacy on local devices. It is a key enabler for implementing artificial intelligence in cyber threat intelligence (CTI) detection scenarios. However, the device-to-device (D2D) communication system of DFL has limitations in communication resources, and data heterogeneity among devices degrades global model performance. This paper proposes a multi-device collaborative optimization method for DFL models (MCO-DFL) based on model quantization and fine-tuning techniques to address these challenges. We use a dynamic differential quantization scheme to reduce the parameter size of local network threat detection models and employ a two-stage model fine-tuning scheme based on knowledge distillation to fine-tune the global model. Extensive experiments show that MCO-DFL reduces communication overhead while maintaining global model performance. Our method outperforms selected SOTA methods on various metrics.

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

MCO-DFL: Multiple Devices Collaborative Optimization for Distributed Federated Learning Based on Cyber Threat Detection Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jianlong Li,
  • Chao Bu,
  • Zishuai Yan

摘要

Distributed Federated Learning (DFL) solves the problems of single-point failure and high latency in traditional federated learning while ensuring data privacy on local devices. It is a key enabler for implementing artificial intelligence in cyber threat intelligence (CTI) detection scenarios. However, the device-to-device (D2D) communication system of DFL has limitations in communication resources, and data heterogeneity among devices degrades global model performance. This paper proposes a multi-device collaborative optimization method for DFL models (MCO-DFL) based on model quantization and fine-tuning techniques to address these challenges. We use a dynamic differential quantization scheme to reduce the parameter size of local network threat detection models and employ a two-stage model fine-tuning scheme based on knowledge distillation to fine-tune the global model. Extensive experiments show that MCO-DFL reduces communication overhead while maintaining global model performance. Our method outperforms selected SOTA methods on various metrics.