Federated & Collaborative Learning with Robustness and Personalization

Martin Jaggi (EPFL)

Federated learning is enabling many promising new applications for machine learning while respecting users' privacy. In this talk, we will discuss the two aspects of 1) robustness to potentially malicious participants and faulty data, and 2) personalization of the trained ML models to each participant, in the realistic setting of heterogeneous data. We will employ tools from stochastic gradient descent algorithms and discuss both theoretical and practical implications in the federated setting, as well as advocating for removing central coordinators to take steps towards fully decentralized collaborative learning.

Coffee Break

Privacy-Preserving Federated Learning

Aurélien Bellet (INRIA)

Federated learning (FL) is a machine learning paradigm where several participants collaboratively train a model while keeping their data decentralized. However, the model parameters or gradients exchanged during the FL training process may leak information about the data. In this talk, I will show how to use the notion of Differential Privacy (DP) to design FL algorithms that provably ensure privacy and confidentiality. In particular, I will present two approaches (one for server-orchestrated FL and one for fully decentralized FL) that nearly match the privacy-utility trade-off of the centralized setting without relying on a trusted curator or complex secure computation primitives.

Related papers:

https://arxiv.org/abs/2006.07218

https://arxiv.org/abs/2012.05326

Fast Distributed Optimization with Asynchrony and Time Delays

Laurent Massoulié (Microsoft Research, Inria Joint Center)

Abstract

The training of models over distributed data calls for distributed optimization schemes. This has motivated research on distributed convex optimization, leading to the identification of lower bounds on convergence speed, and distributed optimization algorithms with convergence speed potentially matching these lower bounds, for a variety of settings.

In this talk we focus on the important setting of asynchronous operation, for which we propose optimization algorithms with optimal speed. We next consider systems with heterogeneous communication and computation delays, for which we propose fast asynchronous algorithms adapted to these heterogeneous delays.

Lunch Break

Lunch will be provided.

Homomorphic Encryption

Pascal Paillier (Zama)

Health data, research and machine learning: what legal framework for professionals?

Yann Favier (University Savoie Mont-Blanc)

Abstract

Health data, qualified as sensitive data, have a special status in the legal regime of data protection defined both in the European framework of the General Data Protection Regulation (GDPR) of April 27, 2016 and by specific laws end others regulations. While public data is being opened up, health data requires an adapted legal treatment that is all the more necessary as artificial intelligence plays a vital role in medicine and clinical research.

Coffee Break