Class-Hidden Client-Side Watermarking in Federated Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Federated learning (FL) offers privacy benefits but faces model leakage risks from participants.
• Protecting intellectual property and model integrity in FL is crucial, especially with unreliable clients.

Purpose of the Study:

• To propose a robust client-side watermarking scheme for federated learning.
• To safeguard model copyright and enhance watermark security against unauthorized access and attacks.

Main Methods:

• Introduced an N+1-class classifier by adding a watermark class to the client model's output layer.
• Trained local models using both watermark and local datasets, storing watermark parameters locally.
• Amplified watermark parameters before server upload and iteratively updated them post-aggregation.

Main Results:

• Achieved 100% watermark detection rates on MNIST, CIFAR-100, and CIFAR-10 datasets using VGG-16 and ResNet-18.
• Demonstrated minimal impact on overall model performance.
• Showcased strong robustness against pruning and fine-tuning attacks.

Conclusions:

• The proposed client-side watermarking scheme effectively protects federated learning models.
• The method enhances model copyright security without compromising performance or robustness.