We study the problem of implementing machine learning algorithms over fully decentralized networks such as networks of mobile phones, sensors, or smart meters. Our algorithms provide a cheap yet usable set of tools that allow data mining over extremely large systems without any investment into specific infrastructure such as cloud services. We have implemented decentralized stochastic gradient descent algorithms for SVM and logistic regression and we have studied privacy issues as well in cryptographic and differential privacy settings. We also proposed several middleware protocols to maximize the performance and efficiency of decentralized protocols.

People

Márk Jelasity (contact), István Hegedűs, Gábor Danner, Árpád Berta