Scholarlink

The multi-fold increase in the number of resource-constrained IoT devices makes the availability of intrusion detection systems (IDS) indispensable, which could be considered to have an optimal combination of computational efficiency along with solid cyber-attack resistance. The proposed privacy-preserving federated learning framework implements lightweight neural networks to permit both local and federated threat detection across IoT ecosystems. The system also implements dynamic feature selection based on Chimp Optimization algorithm-based optimization and model aggregation from privacy-enhanced differential privacy, hence lowering communication overhead by 38% without sacrificing detection accuracy. An additional adversarial robustness module protects against poisoning attacks while ensuring detection accuracy of 95.59%. Layer-wise relevance propagation (LRP) creates an interface for explainable threat classification, ensuring mission-critical transparency. Experiments on benchmark IoT datasets including the MQTTset and CIC-IoV-2024 endorse superior performance, in terms of accuracy, energy efficiency, and communication costs in contrast to others in federated IDS with state-of-the-art performance. This paves the way for scalable, privacy- aware, and interpretable security for next-generation IoT deployments most suited to low-power edge devices with self-imposed stringent resource constraints.

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

Federated Learning-Based Intrusion Detection for Resource-Constrained IoT Devices Using Lightweight Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Navom Saxena,
  • Shraddha Verma,
  • Shubneet,
  • Bhanu Prakash Reddy Rella,
  • Anushka Raj Yadav,
  • Navjot Singh Talwandi

摘要

The multi-fold increase in the number of resource-constrained IoT devices makes the availability of intrusion detection systems (IDS) indispensable, which could be considered to have an optimal combination of computational efficiency along with solid cyber-attack resistance. The proposed privacy-preserving federated learning framework implements lightweight neural networks to permit both local and federated threat detection across IoT ecosystems. The system also implements dynamic feature selection based on Chimp Optimization algorithm-based optimization and model aggregation from privacy-enhanced differential privacy, hence lowering communication overhead by 38% without sacrificing detection accuracy. An additional adversarial robustness module protects against poisoning attacks while ensuring detection accuracy of 95.59%. Layer-wise relevance propagation (LRP) creates an interface for explainable threat classification, ensuring mission-critical transparency. Experiments on benchmark IoT datasets including the MQTTset and CIC-IoV-2024 endorse superior performance, in terms of accuracy, energy efficiency, and communication costs in contrast to others in federated IDS with state-of-the-art performance. This paves the way for scalable, privacy- aware, and interpretable security for next-generation IoT deployments most suited to low-power edge devices with self-imposed stringent resource constraints.