Vers un apprentissage fédéré à large échelle explicable et équitable

Keywords

Federated Learning, big data, explainability, fairness

Subject details

In an increasingly data-driven world, decentralized artificial intelligence (AI) has emerged as a key approach to addressing privacy concerns while adhering to regulations such as the General Data Protection Regulation (GDPR) 1 and the AI Act2 . This inherently minimizes reliance on a single point of data storage and enables decentralized AI to process and use data locally, thereby protecting personal information. Federated learning (FL), one of the main approaches to decentralized AI, enables models to be trained on multiple devices/servers without transferring any data. This approach aligns well with privacy regulations, but it also introduces new challenges in managing client participation (data sources) in a large-scale context. Indeed, the focus on large-scale FL meets the growing demand for AI solutions capable of managing millions of FL clients, which is crucial for the future of AI in many sectors where the use of the Internet of Things is exploding. As the number of FL clients increases, the challenge of managing heterogeneous non-independent and identically distributed data (IID) becomes more pronounced, which can compromise the fairness of the model as well as its explainability. Indeed, fairness in FL can be defined as ensuring that all participating clients benefit in a fair and balanced way from the overall model, according to their contribution and needs. While explainability in FL refers to the ability to understand, interpret and justify the decisions of the global model trained in a decentralized way. In the literature fairness and explainability are often addressed separately. Indeed, both are closely linked, as they both contribute to a more transparent, ethical and reliable system. However, their interaction can be complementary or con:licting, requiring compromise in the design of FL models, especially within a large-scale context. On the one hand, a conflict emerges when improving fairness leads to a more complex model, for example by applying weighting mechanisms to correct biases linked to non IID data. These adjustments may make learning less transparent and more difficult to explain, thus compromising the model's explicability. On the other hand, explicability and fairness are also complementary, as explicability makes it possible to identify inequalities in the distribution of model performance and adjust learning methods to ensure a more fair distribution of benefits. So, while ensuring greater fairness, it is essential not to sacrifice the explainability of the model, thus the need for a compromise: adopting approaches that make learning fair without altering its understanding and transparency too much. The thesis's objective is to create FL models that support ethical principles, with a particular focus on fairness and explainability, even as they scale to accommodate large numbers of clients. Expected contributions - Develop a methodological framework for analyzing the interaction between fairness and explicability in a large-scale federated learning context. - Propose metrics for simultaneously assessing fairness and explainability in a large-scale federated learning system, to better understand their interactions and the necessary trade-offs. - Propose a federated learning approach that guarantees fairness while maintaining robust explicability within a large-scale context. - Validation on open datasets such as MNIST and CIFAR is expected.