PhD Position F/M Fairness in Federated Learning

Key Skills

Algorithms, Experimentation, Fairness learning, Federated learning, Machine learning techniques, Statistical modeling

Job Description

The selected PhD student will be mainly based in Lille in the MAGNET team but will also frequently visit the COMETE team. The main objective of the COMETE team is to develop principled approaches to privacy protection to guide the design of sanitization mechanisms in realistic scenarios. Similarly, the main objective of the MAGNET team is to develop ethically acceptable machine learning algorithms focusing on privacy, federated learning, and fairness and to empower end users of artificial intelligence.

Moreover, the position is part of FedMalin, a large-scale research initiative involving 10 Inria teams across 6 different Inria centers, 22 researchers, 16 PhD students (6 to be hired), 5 postdocs (3 to be hired) and 7 engineers (6 to be hired). The recruited person will have the opportunity to collaborate with other participants of this initiative and take advantage of the scientific emulation it will create.

The PhD candidate will be under the supervision of Aurélien Bellet, who has been working on federated machine learning for several years, Catuscia Palamidessi, who is a specialist in privacy preserving and fair machine learning, and Michaël Perrot, whose main research focus is on the problem of fair machine learning.

Assignment

Machine Learning is now used in digital assistants, for medical diagnosis, for autonomous vehicles, .... Its success can be explained by the good performances of learned models, sometimes reaching human-level capabilities. However, simply being accurate is not sufficient if these models are to be largely deployed and the notion of fairness, and more largely of trustworthiness, has to be considered as soon as humans are involved in the loop. For example, a model used for medical diagnosis or an automated hiring process should not be biased against subgroups of the population. A recent trend in Machine Learning is thus to propose approaches to learn models that are as accurate as possible while satisfying some level of fairness, that is that do not unjustly discriminate against some individuals or subgroups of the population.

Most Machine Learning algorithms were developed in environments where the data can be centralized and easily accessed. However, in many use-cases, data is naturally decentralized and should not be publicly disclosed. For example, medical data is collected and stored by different hospitals or crowdsensed data is generated by personal devices. This raises new challenges and, in this context, Federated Learning emerged as a paradigm where a set of entities with local datasets collaborate to collectively learn models without explicitly sharing their data. The main objective being to reach levels of utility on par with the centralized setting where all the data is owned by a single entity.

While fairness has been widely studied in the centralized setting, the decentralized nature of the data in Federated Learning raises new challenges. For example, the fairness level of the models becomes difficult to measure as each data holding entity only has a partial view of the world. Similarly, as the different entities collaborate, they expect fair rewards that are proportional to their implication. The goal of this PhD is to study fairness in Federated Learning from both a theoretical and an applied point of view. It involves formally understanding the various trade-offs that may arise due to decentralization and proposing sound algorithms able to learn models that are guaranteed to be fair.

Main activities

• Review and follow the existing literature on Fairness and Federated Learning.
• Theoretically and empirically study the Fairness trade-offs inherent to Federated Learning and related to the decentralized nature of the data.
• Propose concrete approaches to measure and enforce various notions of Fairness in Federated Learning, and validate them on real datasets.
• Publish and present results in top machine learning conferences and journals.

Skills

A good candidate will have the following skills:

• A good command of English
• A strong background in mathematics
• A good knowledge of machine learning, statistics and algorithms
• Some experience with implementation and experimentation
• Preferably some knowledge on either fairness or federated learning (or both)

Please follow the instructions given in https://team.inria.fr/magnet/how-to-apply/ to set up your application file.

Benefits package

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking and flexible organization of working hours
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage

Remuneration

1st and 2nd year: 2051 € Gross monthly salary (before taxes) 

3rd year: 2158 € gross monthly salary (before taxes)

General Information

• Theme/Domain: Optimization, machine learning, and statistical methods Statistics (Big data) (BAP E)
• Town/city :Villeneuve d'Ascq
• Inria Center :Centre Inria de l'Université de Lille
• Starting date:2023-04-01
• Duration of contract :3 years
• Deadline to apply:2023-02-02

Contacts

• Inria Team:MAGNET
• Ph.D. Supervisor : Perrot Michael / perrot@inria.fr

The keys to success

A successful candidate will

• Collaborate in the team and where applicable with external researchers and engineers
• Organize work efficiently and make a good balance between the several priorities
• Report regularly