Towards Efficient Backdoor Attacks Against Federated Self-supervised Learning as a Service Through Intra-Union Aggregation

Compared with classical federated learning (FL) with label supervision, federated self-supervised learning (FSSL) has shown soaring performance in representation learning, making it promising to be deployed as a service. Despite the FSSL-as-a-service can collaboratively train a model using unlabeled data across numerous clients, it introduces vulnerabilities to backdoor attacks due to its distributed nature. To complete a backdoor attack in FSSL, several malicious clients upload crafted local models for aggregation so that a target prediction can be activated by a pre-defined trigger pattern. However, the training process of FSSL is beyond the control of malicious clients, which may terminate at any point. Therefore, a backdoor attack with higher efficiency is preferred. In this paper, we propose an efficient backdoor attack on FSSL-as-a-service named United Backdoor Attacks (UBA). Specifically, UBA aggregates models within the union of participating malicious clients to initialize their models in each epoch, which can significantly enhance the attack efficiency. For example, UBA can achieve 20 to 70 times higher attack success rate on CIFAR100 compared to baselines after the same 20 attacking rounds. Furthermore, we demonstrate that UBA can evade the current state-of-the-art defense mechanisms in FL. To mitigate embedded backdoors inside the global model, we also investigate the effectiveness of a general-purpose defense strategy, i.e., fine-tuning, and explore the impacting factors. Our code is available at https://github.com/wsc2000/UBA .