Privacy-aware Optimization and Learning in Multi-agent environments

点击翻译按钮获取中文摘要

In this research project, learning and optimization from privately distributed data sources have been studied. Specifically, algorithm design, security, and performance evaluation has been considered. The traveling salesman problem, reinforcement learning, and pre/post-processing of classification tasks have been focused as target algorithms. In order to securely perform these algorithms with taking privately distributed information, protocol are specifically designed for each algorithm by making use of homomorphic public-key cryptosystem. Furthermore, we proved the security of these protocols. The assessment of the computational efficiency has been performed experimentally.