Federated Learning Backdoor Attack Based on Frequency Domain Injection

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Federated learning (FL) enables collaborative model training without data sharing.
• FL systems are vulnerable to backdoor attacks from malicious clients.
• Existing attacks often use visible triggers and corrupt semantic information.

Purpose of the Study:

• To propose a novel, stealthier backdoor attack for federated learning.
• To overcome the limitations of existing spatial-domain attacks.
• To enhance the effectiveness of backdoor attacks in FL.

Main Methods:

• Developed a frequency-domain injection-based backdoor attack.
• Utilized Fourier transform to mix trigger and clean image in the frequency domain.
• Injected low-frequency trigger information while preserving semantic content.

Main Results:

• The proposed attack is stealthier than existing methods.
• The attack demonstrates higher effectiveness in FL scenarios.
• Experiments were conducted on multiple image classification datasets.

Conclusions:

• Frequency-domain attacks offer a more robust approach to backdoor insertion in FL.
• This method preserves semantic information, making attacks harder to detect.
• The findings highlight new security challenges in distributed machine learning.