Inspecting The Machine Learning Pipeline

Author : Ali Burak Ünal
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (134 download)

DOWNLOAD NOW!

Download or read book Towards a Complete Privacy Preserving Machine Learning Pipeline written by Ali Burak Ünal and published by . This book was released on 2022 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: Machine learning has proven its success on various problems from many different domains. Different machine learning algorithms use different approaches to capture the underlying patterns in the data. Even though the amount varies between the machine learning algorithms, they require sufficient amounts of data to recognize those patterns. One of the easiest ways to meet this need of the machine learning algorithms is to use multiple sources generating the same type of data. Such a solution is feasible considering that the speed of data generation and the number of sources generating these data have been increasing in parallel to the developments in technology. One can easily satisfy the desire of the machine learning algorithms for data using these sources. However, this can cause a privacy leakage. The data generated by these sources may contain sensitive information that can be used for undesirable purposes. Therefore, although the machine learning algorithms demand for data, the sources may not be willing or even allowed to share their data. A similar dilemma occurs when the data owner wants to extract useful information from the data by using machine learning algorithms but it does not have enough computational power or knowledge. In this case, the data source may want to outsource this task to external parties that offer machine learning algorithms as a service. Similarly, in this case, the sensitive information in the data can be the decisive factor for the owner not to choose outsourcing, which then ends up with non-utilized data for the owner. In order to address these kinds of dilemmas and issues, this thesis aims to come up with a complete privacy preserving machine learning pipeline. It introduces several studies that address different phases of the pipeline so that all phases of a machine learning algorithm can be performed privately. One of these phases addressed in this thesis is training of a machine learning algorithm. The privacy preserving training of kernel-based machine learning algorithms are addressed in several different works with different cryptographic techniques, one of which is a our newly developed encryption scheme. The different techniques have different advantages over the others. Furthermore, this thesis introduces our study addressing the testing phase of not only the kernel-based machine learning algorithms but also a special type of recurrent neural network, namely recurrent kernel networks, which is the first study performing such an inference, without compromising privacy. To enable the privacy preserving inference on recurrent kernel networks, this thesis introduces a framework, called CECILIA, with two novel functions, which are the exponential and the inverse square root of the Gram matrix, and efficient versions of the existing functions, which are the multiplexer and the most significant bit. Using this framework and other approaches in the corresponding studies, it is possible to perform privacy preserving inference on various pre-trained machine learning algorithms. Besides the training and testing of machine learning algorithms in a privacy preserving way, this thesis also presents a work that aims to evaluate the performance of machine learning algorithms without sacrificing privacy. This work employs CECILIA to realize the area under curve calculation for two different curve-based evaluations, namely the receiver operating characteristic curve and the precision-recall curve, in a privacy preserving manner. All the proposed approaches are shown to be correct using several machine learning tasks and evaluated for the scalability of the parameters of the corresponding system/algorithm using synthetic data. The results show that the privacy preserving training and testing of kernel-based machine learning algorithms is possible with different settings and the privacy preserving inference on a pre-trained recurrent kernel network is feasible using CECILIA. Additionally, CECILIA also allows the exact area under curve computation to evaluate the performance of a machine learning algorithm without compromising privacy.