Scholarlink

<p>Lane Change Prediction (LCP) is a critical element within Intelligent Transportation Systems (ITS), aimed at improving road safety and streamlining traffic flow. While Deep Learning (DL) models have shown remarkable capabilities in achieving state-of-the-art results in LCP systems, their implementation faces privacy challenges due to the significant data requirements involved. This study endeavors to employ Secure Multiparty Computation (SMPC) technique, enabling multiple parties to collaboratively compute lane change prediction using their private inputs while maintaining confidentiality. Recurrent Neural Network (RNN) models were trained on the HighD dataset. During the inference phase, a Secure Tanh function was applied to uphold privacy in LCP computations. Experimental results indicate that the accuracy of the proposed SMPC-based LCP is nearly equivalent to the traditional methods, while addressing the security concerns. This is supported by Mean Squared Error (MSE) calculations, which fall within the range of 10<sup>–7</sup> for both non-SMPC and SMPC scenarios. Furthermore, we tested the proposed algorithm on homogeneous and heterogeneous setups with different Jetson devices as servers. Experiment results show that due to synchronous nature of the LCP scenario, the speed of inference model is restricted to the less powerful server.</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

Privacy-preserving Lane Change Prediction Using Recurrent Neural Network With Secure Multiparty Computation

  • 期刊论文

  • 发表日期：

    2025年5月3日

查看全文
  • Armin Nejadhossein Qasemabadi,
  • Saeed Mozaffari,
  • Majid Ahmadi,
  • Shahpour Alirezaee

摘要

Lane Change Prediction (LCP) is a critical element within Intelligent Transportation Systems (ITS), aimed at improving road safety and streamlining traffic flow. While Deep Learning (DL) models have shown remarkable capabilities in achieving state-of-the-art results in LCP systems, their implementation faces privacy challenges due to the significant data requirements involved. This study endeavors to employ Secure Multiparty Computation (SMPC) technique, enabling multiple parties to collaboratively compute lane change prediction using their private inputs while maintaining confidentiality. Recurrent Neural Network (RNN) models were trained on the HighD dataset. During the inference phase, a Secure Tanh function was applied to uphold privacy in LCP computations. Experimental results indicate that the accuracy of the proposed SMPC-based LCP is nearly equivalent to the traditional methods, while addressing the security concerns. This is supported by Mean Squared Error (MSE) calculations, which fall within the range of 10–7 for both non-SMPC and SMPC scenarios. Furthermore, we tested the proposed algorithm on homogeneous and heterogeneous setups with different Jetson devices as servers. Experiment results show that due to synchronous nature of the LCP scenario, the speed of inference model is restricted to the less powerful server.