WFB: watermarking-based copyright protection framework for federated learning model via blockchain.

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Blockchain Technology

Background:

• Federated learning (FL) trains models collaboratively without sharing private data, crucial for privacy-sensitive industries.
• FL models face threats like adversarial attacks and ownership generalization, jeopardizing copyright and reliability.
• Current model watermarking for FL ownership verification lacks a credible link between extracted watermarks and verification, hindering closed-loop protection.

Purpose of the Study:

• To propose WFB, a blockchain-empowered watermarking framework for federated model ownership verification.
• To address the credibility issues in current watermark verification schemes for FL.
• To establish a robust closed-loop framework for FL model copyright protection.

Main Methods:

• Developed an improved watermark generation algorithm to enhance watermark credibility.
• Implemented a watermark embedding method within the federated learning process.
• Utilized blockchain technology for secure and credible storage of watermark information.

Main Results:

• WFB demonstrates high fidelity, effectiveness, and robustness in verifying FL model ownership.
• The framework improves the credibility of ownership verification through watermark authenticity.
• Experimental results validate the proposed approach's performance and security enhancements.

Conclusions:

• WFB provides a credible and closed-loop watermarking framework for federated learning.
• The integration of blockchain enhances the security, traceability, and reliability of FL model ownership verification.
• This approach effectively safeguards FL model copyright against various threats.