Journal of AI, Data Science and Cyber Systems

In order to better understand the field of Privacy-Preserving AI and Secure Data Analytics, noting the ‘analytics’ as applicable to any framework of AI which combines the extraction of value from AI and keeping sensitive information protected at all times during its life cycle, is important. This paradigm solves the problem of increasing friction between the needs of data-driven intelligence and the needs of privacy by incorporating the dual principles of privacy and security into data collection, data storage, data processing, and model training. Privacy, security, and the protective principles of data collection differ from the majority of the field, which are focused primarily on control. Conversely, these privacy-preserving principles are focused primarily on the concern that data, even when it is shared and even when it is used for analysis, there has to be some form of protection for the individual, for the sensitive and proprietary claims, and for the data of the owner. Developments include differential privacy, which reduces the chances that individual data points can be gleaned from model outputs; distributed and federated learning, which allow models to be trained, without transferring raw data, at decentralized data repositories; secure multi-party computation and homomorphic encryption, which allow processing of data that remains encrypted; and trusted execution environments, which keep sensitive processing in an isolated space. Secure data analytics employs strong encryption, access control, auditing, and integrity checking to prevent unauthorized data access, loss, or modification. These techniques streamlining safe analytics of sensitive data to meet the legal, ethical, and privacy requirements in the responsible and secure analysis of AI systems in healthcare, finance, public sector, smart cities, etc.