Neural Network Information Leakage through Hidden Learning

Arthur da Cunha, Emanuele Natale, Laurent Viennot

Abstract: We investigate the problem of making an artificial neural network perform hidden computations whose result can be easily retrieved from the network's output. In particular, we consider the following scenario. A user is provided a neural network for a classification task by a third party. The user's input to the network contains sensitive information and the third party can only observe the output of the network. I this work, we provide a simple and efficient training procedure, which we call hidden learning, that produces two networks: (i) one that solves the original classification task with performance near to state of the art; (ii) a second one that takes as input the output of the first, retrieving sensitive information to solve a second classification task with good accuracy. Our result might expose important issues from an information security point of view, as for the use of artificial neural networks in sensible applications.