TL;DR
- 임의의 공격자가 global model의 성능을 악의적으로 성능 저하시킬 수 있음. (model poisoning)
- 효과적인 방어 기법이 없음.
Motivation
Method
Experiments
Image Classification
CIFAR-10, 100 clients, 10 clients / round.
ResNet18 w/ CNN, non-IID, 50000장
How To Backdoor Federated Learning
Category
Federated Learning
Year/Month
2018-07
Status
In progress
Publications
AISTATS
Code