On the Security and Privacy of Collaborative Learning
Event details
| Date | 02.05.2022 |
| Hour | 16:00 › 18:00 |
| Speaker | Mathilde Raynal |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Martin Jaggi
Thesis advisor: Prof. Carmela Troncoso
Co-examiner: Prof. Anne-Marie Kermarrec
Abstract
Collaborative Learning can be used by separate
parties to train a model that will benefit from all private local
datasets. Because parties can be corrupted, a Collaborative
Learning systems need to provide guarantees on the security
of the final model, i.e., on what it will converge to, and on the
privacy of the local datasets. The three papers we survey highlight
different attack points of such systems, exploit them to implement
privacy attacks, and finally discuss potential defenses. All strongly
contribute to the understanding of some of the privacy risks
associated with Collaborative Learning
Background papers
Privacy Risk in Machine Learning: Analyzing the Connection to Overfitting
Understanding Membership Inferences on Well-Generalized Learning Models
Exploiting Unintended Feature Leakage in Collaborative Learning
Exam president: Prof. Martin Jaggi
Thesis advisor: Prof. Carmela Troncoso
Co-examiner: Prof. Anne-Marie Kermarrec
Abstract
Collaborative Learning can be used by separate
parties to train a model that will benefit from all private local
datasets. Because parties can be corrupted, a Collaborative
Learning systems need to provide guarantees on the security
of the final model, i.e., on what it will converge to, and on the
privacy of the local datasets. The three papers we survey highlight
different attack points of such systems, exploit them to implement
privacy attacks, and finally discuss potential defenses. All strongly
contribute to the understanding of some of the privacy risks
associated with Collaborative Learning
Background papers
Privacy Risk in Machine Learning: Analyzing the Connection to Overfitting
Understanding Membership Inferences on Well-Generalized Learning Models
Exploiting Unintended Feature Leakage in Collaborative Learning
Practical information
- General public
- Free