Workshop FL-Day - Decentralized Federated Learning: Approaches and Challenges
Workshop FL-Day - Decentralized Federated Learning: Approaches and Challenges
- Organizers contacts
This scientific day aims at a moment of exchange around the theme of federated and decentralised learning with recent developments from artificial intelligence. Organised by the ADAM team of the DAVID Laboratory as part of the DATAIA Institute of Paris-Saclay.
The day will address, through several presentations, issues related to the theme of "Decentralized Federated Learning", machine learning, decentralized data processing (Edge Computing) or data protection "privacy" in a decentralized context with illustrations in different areas. The presentations will be followed by a round table discussion.
Participants who wish to do so are invited to propose Posters to exhibit their work during the breaks. It is possible to reuse a poster by indicating the event to which it has already been presented.
A buffet lunch and gourmet breaks will be provided throughout the day.
With the participation of the Graduate School of Computing and Digital Sciences.
- 9am - 9:30am : Welcome
- 9:30am - 10:40am : Aurélien Bellet "Better Privacy Guarantees for Decentralized Federated Learning" (1h presentation + 10min questions)
- 10:40am - 11:10am : Coffee break
- 11:10am - 12:20pm : Sonia Ben Mokhtar "Decentralized Learning (as an enabler) for Decentralized Online Services" (1h presentation + 10min questions)
- 12:20pm - 12:50pm Poster pitches
- 12:50pm - 2pm Lunch and posters
- 2pm - 3:10pm : Hakim Hacid "Towards Edge AI: Principles, current state, and perspectives" (1h presentation + 10min questions)
- 3:10pm - 3:40pm : Coffee break
- 3:40pm - 4:50pm : Round table
- 4:50pm - 5pm : Closing
To participate via videoconference, here is the connection link (available from 9:30am).
- Title: Better Privacy Guarantees for Decentralized Federated Learning
- Abstract: Fully decentralised algorithms, in which participants exchange messages peer-to-peer along the edges of a network graph, are becoming increasingly popular in federated learning because of their scalability and efficiency. Intuitively, decentralised algorithms should also offer better privacy guarantees, since nodes only observe messages sent by their neighbours in the graph. But formalizing and quantifying this gain is a challenge: existing results are limited to local differential privacy (LDP) guarantees that neglect the benefits of decentralization. In this talk, I will present appropriate relaxations of differential privacy and show how they can be used to show stronger privacy guarantees for the decentralised DMS, corresponding to the privacy-utility trade-off of the centralised DMS in some contexts. Interestingly, some of these algorithms amplify privacy guarantees as a function of the distance between nodes in the graph, which corresponds well to user expectations of privacy in some use cases.
- Bio: Aurélien Bellet is a researcher at Inria. His current research focuses on the design of privacy-preserving machine learning algorithms in centralised, federated and decentralised environments. Aurélien has been domain chair at major machine learning conferences such as ICML, NeurIPS and AISTATS. He co-organised several international workshops on machine learning and privacy at NeurIPS, CCS and FOCS, as well as the 10th edition of the French multidisciplinary conference on privacy (APVP). He also co-organises FLOW, a webinar on federated learning with over 1000 registered participants.
- Title: Decentralized Learning (as an enabler) for Decentralized Online Services
- Abstract: There is a strong push towards data-driven services at all levels of society and industry. This started with large-scale web applications such as web search engines (e.g. Google, Bing), social networks (e.g. Facebook, Twitter) and recommendation systems (e.g. Amazon, Netflix) and is becoming increasingly ubiquitous with the adoption of wearable devices and the advent of the Internet of Things. All of these services are enabled by the availability of large computing infrastructures, strong advances in artificial intelligence (AI) and in particular machine learning, and the ability to collect and aggregate large amounts of data about users, their environments and their organisations in cloud infrastructures. But while the advances in AI/ML and distributed infrastructures have been significant, the data-driven applications enabled by these advances pose significant challenges to the privacy of their users and can give rise to threats such as censorship, loss of control of personal data and data leakage. More recently, initiatives such as Web 3.0 promise to decentralise online services, at the heart of which AI/ML plays a crucial role in empowering users to regain control of their personal data and preventing a handful of economic actors from over-concentrating decision-making power.
In this talk, I will present the open challenges for designing decentralised learning algorithms that can serve as an enabler for decentralised online services. I will also present recent research work in this context, focusing on the implementation of decentralised recommender systems based on Gossip Learning, and discuss open research directions.
- Bio: Sonia Ben Mokhtar is a CNRS research director at the LIRIS laboratory (UMR 5205) and head of the Distributed Systems and Information Retrieval group (DRIM). She obtained her PhD in 2007 from the University Pierre et Marie Curie before spending two years at the University College London (UK). Her research focuses on the design of resilient and privacy-preserving distributed systems. Sonia has co-authored over 70 papers in peer-reviewed conferences and journals. She has served on the editorial board of IEEE Transactions on Dependable and Secure Computing and has co-chaired major conferences in the field of distributed systems (e.g. ACM Middleware, IEEE DSN). Sonia has been president of ACM SIGOPS France and is currently vice-president of GDR RSD, a national university network of researchers in distributed systems and networks.
- Titre : Towards Edge AI: Principles, current state, and perspectives
- Abstract: The artificial intelligence (AI) community has invested heavily in developing techniques that can digest very large amounts of data to extract valuable information and knowledge. Most techniques, particularly deep learning models, require large amounts of computing and storage power, making them suitable for cloud-based environments. The intelligence is therefore remote from the end user, raising concerns about, for example, data privacy and latency. Edge AI addresses some of the problems inherent in the cloud and focuses on best practices, architectures and processes for extending data AI outside the cloud. Edge AI brings AI closer to the end user and uses, for example, fewer communication resources, as processing is performed directly on the edge device. This presentation will introduce edge AI and give an overview of existing work and potential future contributions.
- Bio: Dr Hakim Hacid is the senior scientist of the AI cross-cutting unit at the Technology Innovation Institute (TII), a leading scientific research centre based in the United Arab Emirates. Prior to joining TII, he was an associate professor at Zayed University and contributed to research in the areas of data analysis, information retrieval and security. He also served as chairman of the Department of Computer Science and Applied Technology. Dr. Hacid has authored over 60 research papers published in leading journals and conferences and holds several industrial patents. His research interests include databases, data mining and analysis, programming, web information systems, natural language processing and security. Hakim Hacid received his PhD in data mining/databases from the University of Lyon, France. He also obtained a double master's degree in computer science (research and professional master's) from the same university.
- Karine BENNIS-ZEITOUNI, DAVID Laboratory, UVSQ
- Zaineb CHELLY DAGDIA-GARCIA, DAVID Laboratory, UVSQ
- Mustapha LEBBAH, DAVID Laboratory, UVSQ