FedIPR: Ownership Verification for Federated Deep Neural Network Models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning models are vulnerable to intellectual property theft and misuse.
• Verifying ownership of federated learning models (FedIPR) is essential for protecting intellectual property rights (IPR).

Purpose of the Study:

• To propose a novel federated deep neural network (FedDNN) ownership verification scheme.
• To enable private watermark embedding and verification for claiming legitimate IPR of FedDNN models.

Main Methods:

• Developed a FedDNN ownership verification scheme enabling private watermark embedding and detection.
• Each client independently verifies watermarks without sharing private data or watermark information.
• Theoretically justified watermark effectiveness through rigorous analysis.

Main Results:

• Demonstrated reliable embedding and detection of watermarks across computer vision and natural language processing tasks.
• Watermarking did not compromise original model performance.
• The scheme proved resilient to various federated training settings and robust against removal attacks.

Conclusions:

• The proposed scheme effectively verifies FedDNN ownership while preserving data and watermark privacy.
• This approach offers a robust solution for protecting IPR in federated learning environments.
• The watermarking technique is practical and effective for securing federated models.