Differentially Private Learning from Label Proportions

Distributed LSTM-Learning from Differentially Private Label Proportions

“…Most algorithms usually focus on one of the two relevant properties, privacy or prediction accuracy. For example, the dp-LLP algorithm [2], introduced by Sachweh et al, is a fully distributed learning algorithm that uses only locally collected data, as well as data from the neighbor nodes, to predict traffic flow. The authors introduced a variant of Label Proportions, originally developed in [3]- [5], extended by Differential Privacy to ensure that data transfer is protected.…”

“…Our approach builds on the experience of the papers presented, and will leverage the data protection properties from [2], as well as comparable performance to [6], [7].…”

“…Because of the lack of a central authority in a fully distributed scenario, encryption does not directly provide a solution. Recent work that combines learning from label proportions with differential privacy (compare next section) [2] is promising, and we therefore utilize Data Aggregation for our approach aswell.…”

“…Therefore [2] extended Data Aggregation (building histograms) by adding Differential Privacy to the histograms. Differential Privacy adds noise to each bin to ensure a specific privacy guarantee.…”

“…Our approach is combining the advantages of DP-LLP [2], and [6] to build a fully decentralized learning algorithm, which ensures privacy and results in good forecasting performance. First, we will describe the general data exchange between the distributed nodes and the integration of Differential Privacy.…”