ACTION TRANSVERSALE 2023 - PROJET FRIENDLY (ACTION TRANSVERSALE 2023 - PROJET FRIENDLY)
Description du projet : FRIENDLY aims to include diversity and inclusion awareness in how data is prepared, engineered, and fragmented by participating nodes and on the models produced and aggregated in the federation by reducing intersectional bias. A dataset might be deemed biased if it does not accurately reflect the population concerning specific sensitive characteristics, often protected or sensitive attributes. The determination of these attributes is contingent on the specific domain; typical examples include race, gender, and age. Bias can be identified by calculating fairness-related metrics that compare the distributions of the groups of interest in various ways or by identifying underrepresented segments of the target population in the dataset (known as representation bias). Bias can arise from the original data collection methods and locations, or it can be introduced, and sometimes even intensified, during the data preparation stages that precede any analytical task. In both scenarios, using data that does not represent a specific population could render the results of the decision system for that population unreliable. Sometimes, even if the outcome is reliable, it might be illegal or undesirable to base any decision on such attributes for reasons specific to the domain. The aim is to propose an approach to reduce the implications of using biased human-related datasets during the pre-processing stage of data analytics pipelines performed in federated learning settings, using some basic functionalities of IBM’s AI Fairness 360 toolkitTutelle gestionnaire : Autre
Dates du contrat : 01/01/2024 - 31/12/2025
Équipe(s) : BD, DRIM
Responsable scientifique LIRIS : Genoveva Vargas-Solar
Partenaires : Université de Gênes, Université des Amériques de Puebla, Université de Milan
URL : http://vargas-solar.com/friendly/