FedPD: Defending federated prototype learning against backdoor attacks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated Learning (FL) enables collaborative model training while preserving data privacy.
• Backdoor attacks pose a significant threat in FL by manipulating model predictions.
• Existing defenses against backdoor attacks in FL incur high computational and communication costs, especially in non-IID settings.

Purpose of the Study:

• To propose a novel defense framework, FedPD, against backdoor attacks in Federated Learning.
• To reduce the overhead associated with defending against malicious clients in FL.
• To maintain high accuracy on the main task while mitigating backdoor threats.

Main Methods:

• The FedPD framework facilitates the exchange of prototypes between servers and clients, rather than model parameters.
• This prototype-based communication inherently prevents the insertion of backdoor channels by malicious participants.
• Prototypes also function as a source of global knowledge to refine local client training.

Main Results:

• FedPD demonstrates superior and consistent defense performance against backdoor attacks compared to existing methods.
• The framework significantly reduces communication overhead by exchanging prototypes.
• FedPD achieved a 90.73% reduction in attack success rate compared to unprotected FedAvg, while maintaining over 90% main task accuracy.

Conclusions:

• FedPD effectively eliminates backdoor attacks at their source by preventing backdoor channel implantation.
• The proposed method is practical for resource-constrained environments and Non-IID data distributions.
• FedPD offers a robust and efficient solution for securing Federated Learning systems against adversarial threats.