Scholarlink

<p>The exponential growth of smart IoT devices demands intelligent, lightweight, and privacy-preserving cyberattack detection mechanisms for resource-constrained environments. While Split Learning (SL) addresses data privacy and distributed computation, traditional SL models are often static and lack adaptability to the heterogeneous and dynamic nature of IoT systems. This work presents PrivEdge-SL, a unified framework that integrates Differential Privacy (DP) with an optimization pipeline comprising pruning, quantization, and TensorFlow Lite conversion to generate multiple efficient SL model variants. DP is applied to client-side activations to ensure quantifiable privacy–utility trade-offs with minimal communication overhead. The key contribution of this work lies in the integration of a tree-based meta-inference selector that dynamically selects the most suitable SL variant based on device capacity, attack type, and privacy level. Experimental results indicate that PrivEdge-SL reduces memory consumption by up to 72% and inference latency by up to 68% through dynamic selection of optimized model variants, while maintaining classification accuracy within 1–2% of the standard baseline model. This framework contributes a significant step toward scalable and privacy-preserving solutions for secure and real-time IoT intelligence.</p>

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

A privacy preserving split learning framework with adaptive inference selection for IoT security

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月16日

查看全文
  • V. Santhosh Kumar,
  • Dhiraj Sunehra

摘要

The exponential growth of smart IoT devices demands intelligent, lightweight, and privacy-preserving cyberattack detection mechanisms for resource-constrained environments. While Split Learning (SL) addresses data privacy and distributed computation, traditional SL models are often static and lack adaptability to the heterogeneous and dynamic nature of IoT systems. This work presents PrivEdge-SL, a unified framework that integrates Differential Privacy (DP) with an optimization pipeline comprising pruning, quantization, and TensorFlow Lite conversion to generate multiple efficient SL model variants. DP is applied to client-side activations to ensure quantifiable privacy–utility trade-offs with minimal communication overhead. The key contribution of this work lies in the integration of a tree-based meta-inference selector that dynamically selects the most suitable SL variant based on device capacity, attack type, and privacy level. Experimental results indicate that PrivEdge-SL reduces memory consumption by up to 72% and inference latency by up to 68% through dynamic selection of optimized model variants, while maintaining classification accuracy within 1–2% of the standard baseline model. This framework contributes a significant step toward scalable and privacy-preserving solutions for secure and real-time IoT intelligence.