METHODS FOR PRIVACY-PRESERVING IN MACHINE LEARNING

Abstract

Data security and confidentiality are the biggest problems today, as much of the information is stored electronically and transmitted through a variety of devices (smartphones, computers) that have become widespread in public life. This is confirmed by the strengthening of legislation aimed at ensuring data protection. In particular, in 2016, the European Union adopted the General Data Protection Regulation (GDPR), and the California Consumer Privacy Act (CCPA) was adopted in California in 2018. The legal acts mentioned above encourage companies, developers, and researchers to develop information systems that are confidential by design (implementing the "Privacy by Design" approach). Data privacy is also in the central place during the construction of data analysis and artificial intelligence systems. First, such systems penetrate deeper into many areas of human activity every year, such as e-commerce (product recommendations, online assistants), human resource management (candidate resume analysis), social sphere (anti-spam, removal of offensive content), medicine, the gaming industry, and even politics. Also, an important element for building a reliable and accurate system is the availability of sufficient data for training. However, not all collected data can be used for training in decisions using artificial intelligence, as the data may contain a significant amount of private information: secret (eg, financial, military data), confidential (identifying data: passport data, registration number) taxpayer) or sensitive (medical data containing patient diagnoses). Under such conditions, finding a dataset to build an artificial intelligence system is an important task.

The paper presents the results of the analysis of attacks on machine learning systems, as well as countermeasures to preserve the privacy of private data sets: anonymization, federated learning, homomorphic encryption, secure multilateral computing, and differential privacy.