ENHANCING THE PRIVACY OF FEDERATED LEARNING THROUGH DATA SYNTHESIS

M Yashwanth, Sai Anuroop Kesanapalli, Gaurav Kumar Nayak, Anirban Chakraborty, Yogesh Simmhan

Abstract:

Federated Learning (FL) is a distributed machine learning architecture where edge devices collaboratively learn the shared model, while the training data is securely held at the edge devices. FL promises a way forward to preserving data privacy by sending model updates in the form of gradients or the weights themselves. However, these updates still contain the essence of the original training data and can be
reconstructed using gradient-based attacks. To overcome this, we propose a novel Privacy-Preserving Federated Learning algorithm (PPFed) wherein we generate a condensed dataset from the original training data at each edge device. The client’s then train their local models on the condensed dataset which is then broadcasted to the server, followed by regular federated averaging. Our method provides privacy by being robust against gradient-based attacks, which holds across different benchmark datasets and CNN based architectures

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)