Federated Learning with Real-world Datasets: Compliance with the Privacy Act

Federated Learning with Real-world Datasets: Compliance with the Privacy Act

Federated learning has been demonstrated to face challenges when applied into real-world environment whereas data are prone to be not independent and identically distributed (non-IID) and imbalanced. Moreover, despite its privacy-aware paradigm, standard federated learning is vulnerable to privacy attacks where the model parameters exchanged among participants may contain some information that can be exploited to reveal various sensitive information related to participant training data or their properties. With privacy-preserving federated learning proposed to tackle privacy attacks, the challenge of achieving a trade-off between privacy and accuracy has been a concern, particularly, in a real-world privacy-preserving federated learning setting. This paper reviews state-of-the-art federated learning schemes, including both standard and privacy-preserving federated learning schemes, with a focus on those tested with real-world data. Furthermore, we noticed that federated learning-based applications are not naturally compliant with government defined data protection regulations such as the Privacy Act (i.e., the 13 Australian Privacy Principles), which is similar to other machine learning-based applications. In this paper, we investigate the compliance of federated learning against the Privacy Act, including a summary of challenges encountered in terms of complying some of the Australian Privacy Principles (APPs). Recommendations including remedies are suggested for enhancing an FL scheme for complying with all the 13 APPs, that is, the entire Privacy Act.

Files associated with this item cannot be displayed due to copyright restrictions.