Clean-Label Backdoor Attack In Federated Learning

Federated learning is a distributed machine learning approach that enables different clients to participate in training a global model without sharing their data directly. Due to the inherent nature of federated learning, it is susceptible to backdoor attacks. These attacks aim to manipulate the model's output behavior to produce the desired result when it encounters images that have been carefully crafted by attackers. Currently, the majority of backdoor attacks in federated learning are dirty-label attacks, which involve modifying the labels of poisoned samples. However, there is a lack of research on clean-label backdoor attacks in federated learning that do not require label flipping. In this paper, we propose a method for a clean-label backdoor attack in federated learning, called FCLB. Our approach involves utilizing a surrogate model to train and generate optimized triggers. These triggers are then blended with benign images to create the poisoned dataset for malicious clients. During training, we use a gradient ascent algorithm to unlearn a small number of samples from the target label, ensuring the effectiveness of the clean-label backdoor attack. We conducted comprehensive experiments, and the results demonstrate that our clean-label backdoor successfully implants a backdoor into the global model, proving the effectiveness of our attack.