Byzantine Resilience and Privacy in Machine Learning
Event details
| Date | 10.05.2022 |
| Hour | 14:00 › 16:00 |
| Speaker | Youssef Allouah |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Jean-Pierre Hubaux
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Martin Jaggi
Abstract
In this proposal, we analyze the problem of combining two crucial security aspects of distributed machine learning. The first one is Byzantine resilience, that is robustness to faulty or adversarial nodes during training. The second on is differential privacy, a strong standard for guaranteeing the privacy of databases in machine learning. For this, we discuss three existing works. The first two works are separately tackling Byzantine resilience and differential privacy respectively. The third work studies the combination of a specific type of Byzantine resilience with differential privacy. Finally, in light of this analysis, we outline our research proposal.
Background papers
Exam president: Prof. Jean-Pierre Hubaux
Thesis advisor: Prof. Rachid Guerraoui
Co-examiner: Prof. Martin Jaggi
Abstract
In this proposal, we analyze the problem of combining two crucial security aspects of distributed machine learning. The first one is Byzantine resilience, that is robustness to faulty or adversarial nodes during training. The second on is differential privacy, a strong standard for guaranteeing the privacy of databases in machine learning. For this, we discuss three existing works. The first two works are separately tackling Byzantine resilience and differential privacy respectively. The third work studies the combination of a specific type of Byzantine resilience with differential privacy. Finally, in light of this analysis, we outline our research proposal.
Background papers
- Deep Learning with Differential Privacy (https://arxiv.org/abs/1607.00133)
- Byzantine-Robust Distributed Learning: Towards Optimal Statistical Rates (https://arxiv.org/abs/1803.01498)
- Differential Privacy and Byzantine Resilience in SGD: Do They Add Up? (https://arxiv.org/abs/2102.08166)
Practical information
- General public
- Free