Targeted Clean-Label Poisoning Attacks on Federated Learning

Abstract

Federated Learning (FL) has become one of the most extensively utilized distributed training approaches since it allows users to access large datasets without really sharing them. Only the updated model parameters are exchanged with the central server after the model has been trained locally on the devices holding the data. Because of the distributed nature of the FL technique, there is a possibility of adversarial attacks that aim to manipulate the behavior of the model. This paper explores targeted clean-label attack in which adversaries inject poisoned images into compromised clients’ dataset to alter the behaviour of the model on a specific target image at test time. The standard CIFAR10 dataset is used in this study to conduct various experiments and manipulate the image classifier. This study discovered that the behavior of a FL model can be altered maliciously towards a specific target image without significantly affecting the model’s overall accuracy. In addition, the attack’s impact grows in direct proportion to the number of injected poisonous images and malicious client (i.e. controlled by adversaries) participating in the FL process.