Lynda FERRAGUIG

Stéphane PerrinDoctoral student, Polytech Annecy-Chambéry.

Doctoral student in computer science

 

 

  • Contact

Email/ E-mail: lynda.ferraguig@univ-smb.fr / gl_ferraguig@esi.dz

Phone number/ Téléphone : +33 (0)780863227

Fax: +33 (0)4 50 09 65 59

Office : A221

Address : LISTIC - Polytech Annecy-Chambéry, BP 80439 - Annecy le Vieux, 74944 ANNECY Cedex, France

  •  Thesis Title/ titre Thèse

Bias mitigation for collaborative and ethical learning on dynamic data

Bias reduction for collaborative and ethical learning on dynamic data

  • Reaserch team/ Group

AFute/ Look

  • Abstract/ Summary:

Machine learning is increasingly used in decision making processes and allows us to solve increasingly complex problems. However, this approach increases the risk of discriminating against certain populations of data, whether they are related to physical systems or to people. This risk is notably linked to the biases introduced in the learning process. Federated learning, a new machine learning paradigm that is gaining ground in response to the issues of data confidentiality and decentralization of calculations, is also concerned. This collaborative approach keeps the data close to its source and makes the management of bias more complex. Indeed, the confidentiality constraints and, depending on the case, the privacy protection imposed in this approach do not allow the use of classical bias mitigation techniques. Moreover, one of the purposes of federated learning is to build models adapted to hierarchically organized populations in order to generate intermediate models adapted to groups of different populations. Fundamental questions then arise on how to create and control the temporal evolution of this hierarchy while preserving data confidentiality and reducing bias. The state of the art does not currently report any work on this general framework or is limited to a global non-hierarchical approach. The objective of the thesis is then to propose methods for detecting and eliminating both global and sub-population biases by taking into account the dynamic aspect of the data and the privacy constraints by relying on realistic case studies from the state of the art and from projects within the LISTIC.

 

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Machine learning is increasingly used in decision making processes and allows us to solve increasingly complex problems. However, this approach increases the risk of discriminating against certain populations of data, whether they are related to physical systems or to people. This risk is notably linked to the biases introduced in the learning process. Federated learning, a new machine learning paradigm that is gaining ground in response to the issues of data confidentiality and decentralization of calculations, is also concerned. This collaborative approach keeps the data close to its source and makes the management of bias more complex. Indeed, the confidentiality constraints and, depending on the case, the privacy protection imposed in this approach do not allow the use of classical bias mitigation techniques. Moreover, one of the purposes of federated learning is to build models adapted to hierarchically organized populations in order to generate intermediate models adapted to groups of different populations. Fundamental questions then arise on how to create and control the temporal evolution of this hierarchy while preserving data confidentiality and reducing bias. The state of the art does not currently report any work on this general framework or is limited to a global non-hierarchical approach. The objective of the thesis is then to propose methods for detecting and eliminating both global and sub-population biases by taking into account the dynamic aspect of the data and the privacy constraints by relying on realistic case studies from the state of the art and from projects within the LISTIC.

  • Key-words/ Mots-clefs :

  1. Bias
  2. Collaborative learning
  3. Ethical learning
  4. Dynamic data

 

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

  1. Bias
  2. Collaborative learning
  3. Ethical learning
  4. Dynamic data

  • Publications

https://hal.science/hal-03343288/document

  • Phd supervisor/ Encadrant 

Pr Benoit Alexandre

  • Phd co-supervisor/ Co-supervisor

Dr. Loukil Faiza

Start of the thesis/Début de la thèse

15/11/22

  • Doctoral school/ Ecole doctorale

EIS