Information Leaks in Federated Learning

With the surge in data collection and analytics, concerns are raised with regards to the privacy of the individuals represented by the data. In settings where the data is distributed over several data holders, federated learning offers an alternative to learn from the data without the need to centralize it in the first place. This is achieved by exchanging only model parameters learned locally at each data holder. This greatly limits the amount of data to be transferred, reduces the impact of data breaches, and helps to preserve the individual’s privacy. Federated learning thus becomes a viable alternative in IoT and Edge Computing settings, especially if the data collected is sensitive.

However, risks for data or information leaks still persist, if information can be inferred from the models exchanged. This can e.g. be in the form of membership inference attacks. In this paper, we investigate how successful such attacks are in the setting of sequential federated learning. The cyclic nature of model learning and exchange might enable attackers with more information to observe the dynamics of the learning process, and thus perform a more powerful attack.