Ensuring privacy of data in Machine Learning

Machine Learning is one of the most popular advancements in technology which is being widely used in various domains, including healthcare, avionics, automotive, business, education, etc. A Machine learning approach works by learning the required knowledge from the data to be supplied by a client system. In a typical scenario, a client entrusts a third-party agency to develop a machine learning application and shares the data to the developer to enable the development of a Machine learning application. This is badly affecting the privacy of data as a third-party is getting access to sensitive data from a client system. Therefore, an effective data encoding technique is required to ensure the privacy of sensitive data while enabling a third-party agency to develop a Machine learning application on the encoded data. Existing data masking/encoding techniques such as pseudonymization, anonymization and substitution are badly affecting the machine learning process as the modification they do for masking data is preventing a machine learning approach to learn the required knowledge. Another approach known in the literature is Fully Homomorphic Encryption. But, there is no tool available based on this technique which enables a client system to mask sensitive data before giving it to a third-party Machine Learning developer. We failed to obtain the expected outcome from an off-the-shelf Machine Learning classifier when we tried to classify a benchmark dataset masked using an available implementation of Fully Homomorphic Encryption. Since enough details about the implementation is not available, we could not find and fix the issues which is causing this problem. We propose a technique based on which we implemented a tool which provides an easy-to-use Graphical User Interface enabling the client to mask sensitive data before giving it to a third-party agency for developing a Machine Learning Application. Our tool enables a client to mask sensitive data through few mouse clicks. We mask sensitive data by applying a modified version of the RSA encryption system. We have applied our tool on three benchmark datasets and applied an off-the-shelf Machine Learning classifier on the datasets masked by our tool and found that the results obtained on the masked datasets are as good as the results obtained on the unmasked datasets.