Scholarlink

The Internet of Things (IoT) is highly vulnerable to cyber threats due to distributed device deployment and sensitive data exposure. Decentralized IoT systems further suffer from intermittent client connectivity and severely imbalanced attack data. This paper proposes DT-SFL, an adaptive synthetic data generation framework in Digital Twin-enabled Split Federated Learning (DT-SFL) for imbalanced IoT security. Local digital twins generate synthetic samples to address data imbalance during on-device training. The split architecture enhances privacy by transmitting intermediate features rather than raw data or full gradients. Dynamic client interactions and dropout are handled through adaptive digital twin adjustments. The methodology is evaluated under 5 clients, achieving detection accuracies of 0.9996 on CICIDS2017, 0.9224 on CICIIoT2025, and 0.9991 on N-BaIoT. These results enable collaborative, privacy-preserving intrusion detection across distributed IoT ecosystems.

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

DT-SFL: Adaptive Synthetic Data Generation in Digital Twin-Enabled Split Federated Learning for Imbalanced Data in IoT Security

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Farhan Ullah,
  • Leonardo Mostarda,
  • Diletta Cacciagrano,
  • Ghassen Ben Brahim,
  • Khalid Mahmood,
  • Yue Zhao

摘要

The Internet of Things (IoT) is highly vulnerable to cyber threats due to distributed device deployment and sensitive data exposure. Decentralized IoT systems further suffer from intermittent client connectivity and severely imbalanced attack data. This paper proposes DT-SFL, an adaptive synthetic data generation framework in Digital Twin-enabled Split Federated Learning (DT-SFL) for imbalanced IoT security. Local digital twins generate synthetic samples to address data imbalance during on-device training. The split architecture enhances privacy by transmitting intermediate features rather than raw data or full gradients. Dynamic client interactions and dropout are handled through adaptive digital twin adjustments. The methodology is evaluated under 5 clients, achieving detection accuracies of 0.9996 on CICIDS2017, 0.9224 on CICIIoT2025, and 0.9991 on N-BaIoT. These results enable collaborative, privacy-preserving intrusion detection across distributed IoT ecosystems.