A distributed privacy-preserving regularization network committee machine of isolated Peer classifiers for P2P data mining

For distributed data mining in peer-to-peer systems this work describes a completely asynchronous, scalable and privacy-preserving committee machine. Regularization neural networks are used for all the Peer classifiers and the combiner committee in an embedded architecture. The proposed method builds the committee machine using the large amounts of training data distributed over the peers, without moving the data, and with little centralized coordination. At the end of the training phase no Peer will know anything else besides its own local data. This privacy-preserving obligation is a challenging problem for trainable combiners but is crucial in real world applications. Only classifiers are transmitted to other peers to validate their data and send back average accuracy rates in a classical asynchronous peer-to-peer execution cycle. Here the validation set for one classifier becomes the training set of the other and vice versa. From this entirely distributed and privacy-preserving mutual validation a coarse-grained asymmetric mutual validation matrix can be formed to map all Peer members. We demonstrate here that it is possible to exploit this matrix to efficiently train another regularization network as the combiner committee machine.