Differential Privacy and Neural Networks: A Preliminary Analysis

The soaring amount of data coming from a variety of sources including social networks and mobile devices opens up new perspectives while at the same time posing new challenges. On one hand, AI-systems like Neural Networks paved the way toward new applications ranging from self-driving cars to text understanding. On the other hand, the management and analysis of data that fed these applications raises con- cerns about the privacy of data contributors. One robust (from the mathematical point of view) privacy definition is that of Differential Privacy (DP). The peculiarity of DP-based algorithms is that they do not work on anonymized versions of the data; they add a calibrated amount of noise before releasing the results, instead. The goals of this paper are: to give an overview on recent research results marrying DP and neural net- works; to present a blueprint for differentially private neural networks; and, to discuss our findings and point out new research challenges.