Strategyproof and Byzantine Resilient Personalized Collaborative Learning
Event details
| Date | 13.07.2021 |
| Hour | 14:00 › 16:00 |
| Speaker | Sadegh Farhadkhani |
| Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Emmanuel Abbé
thesis advisor: Prof. Rachid Guerraoui
co-examiner: Prof. Nicolas Flammarion
Abstract
Collaborative learning is a proposed framework that enables different nodes to learn from each other's data. In modern applications, this is very appealing, as the dataset of a single node is often too small to achieve acceptable performances. However, in this setting, Byzantine or strategic nodes may deviate from the protocol in order to sabotage the learning procedure or bias the models learned by other nodes in their favor. The latter issue can particularly be the case in recommender systems, where there are significant social and economic incentives to promote certain viewpoints and products. Therefore, we aim to design a personalized collaborative learning framework that is both Byzantine resilient and strategyproof.
In this report, first, we introduce a new formulation for federated learning that enables each node to learn a personalized model while using other nodes' data. Second, we discuss a Byzantine resilient distributed learning algorithm that is robust to malicious workers sending forged gradient estimates. Third, we introduce a linear regression framework that incentivizes honesty for the strategic nodes that want to pull the regression model towards their own desired points. Finally, we discuss our research plans.
Background papers
exam president: Prof. Emmanuel Abbé
thesis advisor: Prof. Rachid Guerraoui
co-examiner: Prof. Nicolas Flammarion
Abstract
Collaborative learning is a proposed framework that enables different nodes to learn from each other's data. In modern applications, this is very appealing, as the dataset of a single node is often too small to achieve acceptable performances. However, in this setting, Byzantine or strategic nodes may deviate from the protocol in order to sabotage the learning procedure or bias the models learned by other nodes in their favor. The latter issue can particularly be the case in recommender systems, where there are significant social and economic incentives to promote certain viewpoints and products. Therefore, we aim to design a personalized collaborative learning framework that is both Byzantine resilient and strategyproof.
In this report, first, we introduce a new formulation for federated learning that enables each node to learn a personalized model while using other nodes' data. Second, we discuss a Byzantine resilient distributed learning algorithm that is robust to malicious workers sending forged gradient estimates. Third, we introduce a linear regression framework that incentivizes honesty for the strategic nodes that want to pull the regression model towards their own desired points. Finally, we discuss our research plans.
Background papers
Practical information
- General public
- Free