Secure SVM Training Over Vertically-Partitioned Datasets Using Consortium Blockchain for Vehicular Social Networks

• Published in: IEEE transactions on vehicular technology Vol. 69; no. 6; pp. 5773 - 5783
• Main Authors: Shen, Meng, Zhang, Jie, Zhu, Liehuang, Xu, Ke, Tang, Xiangyun
• Format: Journal Article
• Language: English
• Published: New York IEEE 01-06-2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Blockchain; Classifiers; Consortia; consortium blockchain; Cryptography; Data retrieval; Datasets; Encryption; Machine learning; Privacy; Privacy Preserving; Social networks; support vector machine; Support vector machines; Training; Trusted third parties; Vehicular Social Networks
• ISSN: 0018-9545; 1939-9359
• DOI: 10.1109/TVT.2019.2957425

Machine learning (ML) techniques are expected to be used for specific applications in Vehicular Social Networks (VSNs). Support vector machine (SVM) is one of the typical ML methods and widely used for its high efficiency. Due to the limitation of data sources, the data collected by different entities usually contain attributes that are quite different. However, in some real-world scenarios, when training an SVM classifier, many entities face the same problem that they are lacking in data with adequate attributes. Thus multiple entities are required to share data to combine a dataset with diverse attributes and then jointly train a comprehensive classifier. However, data privacy concerns are raised because of data sharing. To sovle the problem, we propose a privacy-preserving SVM classifier training scheme over vertically-partitioned datasets posessed by multiple data providers. In our scheme, we utilize consortium blockchain and threshold homomorphic cryptosystem to establish a secure SVM classifier training platform without a trusted third-party. We keep lots of training operations locally over original data and necessary interactions between participants are protected by the threshold Paillier and consortium blockchain. Security analysis proves that our scheme can preserve the privacy of the original data and the training intermediate values. Extensive experiments indicate that our scheme has high efficiency and no accuracy loss.