GuardianML: Anatomy of Privacy-Preserving Machine Learning Techniques and Frameworks

Machine learning has become integral to our lives, finding applications in nearly every aspect of our daily routines. However, using personal information in machine learning applications has raised concerns about user data privacy and security. As concerns about data privacy grow, algorithms and techniques for achieving robust privacy-preserving machine learning (PPML) have become a pressing technical challenge. PPML aims to safeguard the confidentiality of both data and models and ensure that sensitive information remains protected during training and inference processes. Various techniques, protocols, libraries, and frameworks have been proposed for PPML, but choosing the right combination along with the appropriate algorithmic or system parameters for a specific deployment instance can be very difficult. In this work, we introduce GuardianML, an open-source recommendation system for selecting the correct parameters and suitable framework for specific use cases of PPML. GuardianML allows users to search through a wide range of PPML frameworks, techniques, protocols, libraries, and more based on a set of objectives. GuardianML filters potential frameworks based on user-defined criteria, such as the number of parties involved in multi-party computation or the need to minimize communication costs in homomorphic encryption scenarios. The system’s recommendations and optimizations are formulated as a maximization problem using linear integer programming to identify the most suitable solution for various use cases. Moreover, this work thoroughly analyzes and presents seventy relevant frameworks in the system’s database. Additionally, we offer an open-source repository containing practical examples and documentation for some of the frameworks.