Towards Bias Mitigation in Federated Learning

Yasmine Djebrouni (1, 2, 3)

1 LIG - Laboratoire d'Informatique de Grenoble
2 LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information
3 DRIM - Distribution, Recherche d'Information et Mobilité

Federated Learning (FL) provides better user data privacy, while allowing users to collaboratively solve a machine learning problem. However, FL can exacerbate the problem of model bias and unfairness, thus, resulting in segregative or sexist models. The objective of our PhD work is threefold: (i) Characterize the actual impact of FL settings on bias; (ii) Propose novel FL selection and aggregation methods for bias mitigation; (iii) Take into account antagonistic aspects such privacy, bias and robustness in FL.

Domaines: Intelligence artificielle [cs.AI]

Yasmine Djebrouni. Towards Bias Mitigation in Federated Learning. 16th EuroSys Doctoral Workshop, Apr 2022, Rennes, France. ⟨hal-03639179⟩