Privacy-Preserving Learning of Random Forests Without Revealing the Trees.

The paper presents a method for the privacy-preserving learning of random forests from private data of three parties, where not even the decision trees, i.e., neither the tree structures nor their parameters (the annotations of attributes and attribute values), are disclosed to any of the parties. To make this practical for realistically size data, a custom protocol is needed for the private comparison of two numbers, such that the numbers themselves are only available in shares and are not known to either party. Experiments with five datasets indicate that the overall protocol matches classical random forests in accuracy and can handle datasets of realistic size.