Short-Term Scientific Mission

Exploiting Evolutionary Algorithms for Synthetic Data Generation

Main theme: Single- and Multiobjective Optimisation
Grantee: Davide Farinati, Universidade Nova de Lisboa, Lisboa, Portugal
Host: Eric Medvet, University of Trieste, Trieste, Italy

Start date: 2025-04-22
End date: 2025-06-20
Awarded: 2025-03-07
Report approved: 2025-06-30

Description

This STSM focuses on applying Grammatical Evolution (GE) to develop a privacy-preserving synthetic data generation framework. The main goal is to generate datasets that maintain critical information for machine learning tasks while safeguarding sensitive data. Planned activities include designing grammars to represent data generation rules, developing a GE-based optimization model, and conducting benchmark experiments to evaluate data utility and privacy protection. This work will contribute to ROAR-NET’s objectives by demonstrating the practical application of evolutionary algorithms in real-world data privacy challenges and fostering cross-disciplinary knowledge exchange between optimization and data privacy domains.

Achievements

During the STSM, Multi-Tree Genetic Programming (MTGP) was employed to develop a method for privacy-preserving synthetic data generation. The work focused on formulating the task as a bi-objective optimization problem balancing data utility and privacy. A software framework was implemented and validated on four benchmark datasets, demonstrating its ability to replicate machine learning patterns without disclosing sensitive information. The code was released as open-source, and a scientific publication is in preparation. The STSM fostered collaboration between institutions and contributed to the goals of ROAR-NET by advancing randomized optimization techniques for privacy-aware machine learning.