Online Prediction with Privacy

In this paper, we consider online prediction from expert advice in a situation where each expert observes its own loss at each time while the loss cannot be disclosed to others for reasons of privacy or confidentiality preservation. Our secure exponential weighting scheme enables exploitation of such private loss values by making use of cryptographic tools. We proved that the regret bound of the secure exponential weighting is the same or almost the same with the well-known exponential weighting scheme in the full information model. In addition, we prove theoretically that the secure exponential weighting is privacy-preserving in the sense of secure function evaluation.

• Type: Journal
• Year: 2010
• Where: ICML
• Authors: Jun Sakuma, Hiromi Arai
• Keywords: Exponential Weighting Scheme, ICML 2010, Machine Learning, Secure Exponential Weighting, Well-known Exponential Weighting