Matrix Commitment-based Ownership Verification for Distributed Machine Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cryptography

Background:

• Distributed Machine Learning (DML) enables parallel training but faces challenges in determining model ownership and verifying client contributions.
• Existing methods for ownership verification can be computationally intensive or lack robust security against malicious clients.

Purpose of the Study:

• To propose a novel scheme, MAMMON (Matrix Commitment-based DML Ownership Verification), for verifying client contributions and establishing partial ownership in DML.
• To ensure computational integrity and correctness with limited computational cost, addressing limitations of existing approaches.

Main Methods:

• Developed a Matrix Commitment-based scheme (MAMMON) utilizing a multi-linear tree structure to reduce proof update costs.
• Standardized the training process to avoid complex arithmetic circuit computations.
• Implemented watermarking of weight proofs with client identity private keys to prevent tampering.

Main Results:

• MAMMON provides a concise proof for computational integrity and correctness within limited computing costs.
• The scheme effectively reduces proof update costs compared to SNARK-based approaches.
• Experimental results demonstrate MAMMON's superior performance in preserving the computational integrity of DML training.

Conclusions:

• MAMMON offers an efficient and secure solution for DML ownership verification, fairly establishing partial ownership based on client contributions.
• The proposed method enhances the trustworthiness of DML systems by preventing malicious data forging and ensuring computational integrity.