Privacy preserving growing neural gas over arbitrarily partitioned data

The growing neural gas is a well-known neural network model in unsupervised learning missions, such as vector quantization and clustering. With the rampant growth of data and the rapid development of database technology, sometimes multiple data providers may want to cooperate with each other to get better data analysis results by sharing their data. So concern of privacy preserving arises because each party may not be willing to let the others know their valuable data. In this paper, we present a privacy preserving protocol for the multiparty training of growing neural gas while the data are arbitrarily partitioned over different parties. The main idea of our algorithms is that for each input instance of which the attributes are shared by two parties, we randomly split every party׳s own attribute values into two parts, one for private use and the other one for the public exchange with the other party. In this way, each party will have a share of each attribute, no matter which party it is originally owned by. We also prove that our algorithm provides strong privacy guarantees under a standard security model. Besides, the experiments on real world datasets show that our algorithm is also efficient.