TY - GEN
T1 - Transferable Unique Copyright Across AI Model Trading
T2 - 22nd IEEE International Conference on Software Quality, Reliability and Security Companion, QRS-C 2022
AU - Fan, Yixin
AU - Hao, Guozhi
AU - Wu, Jun
N1 - Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - Currently, Machine Learning as a Service (MLaaS) greatly benefits artificial intelligence (AI) model trading. However, threats such as model piracy and patent grabbing devastatingly violate the copyright of AI models. Current invasive copyright protection solutions mainly rely on watermarking to embed specific information into AI models, which inevitably decreases the accuracy. While non-invasive schemes, such as adversarial samples, cannot guarantee uniqueness as the adversarial sample generation algorithm would be known to all traders, and thus need to be changed after trading. To enable the ownership information transferable across AI model trading, we propose a blockchain-driven Non-Fungible Token (NFT) approach for trading-oriented AI model copyright protection. We design a mapping mechanism from AI models parameters to NFTs which can identify uniqueness and ownership of AI models across trading. Besides, a reputation-based rewards and penalties scheme is proposed to prevent NFT piracy. Lastly, the evaluation verifies the applicability of our approach.
AB - Currently, Machine Learning as a Service (MLaaS) greatly benefits artificial intelligence (AI) model trading. However, threats such as model piracy and patent grabbing devastatingly violate the copyright of AI models. Current invasive copyright protection solutions mainly rely on watermarking to embed specific information into AI models, which inevitably decreases the accuracy. While non-invasive schemes, such as adversarial samples, cannot guarantee uniqueness as the adversarial sample generation algorithm would be known to all traders, and thus need to be changed after trading. To enable the ownership information transferable across AI model trading, we propose a blockchain-driven Non-Fungible Token (NFT) approach for trading-oriented AI model copyright protection. We design a mapping mechanism from AI models parameters to NFTs which can identify uniqueness and ownership of AI models across trading. Besides, a reputation-based rewards and penalties scheme is proposed to prevent NFT piracy. Lastly, the evaluation verifies the applicability of our approach.
KW - AI model trading
KW - Blockchain
KW - NFT
KW - copyright protection
UR - http://www.scopus.com/inward/record.url?scp=85152623105&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85152623105&partnerID=8YFLogxK
U2 - 10.1109/QRS-C57518.2022.00023
DO - 10.1109/QRS-C57518.2022.00023
M3 - Conference contribution
AN - SCOPUS:85152623105
T3 - Proceedings - 2022 IEEE 22nd International Conference on Software Quality, Reliability and Security Companion, QRS-C 2022
SP - 102
EP - 105
BT - Proceedings - 2022 IEEE 22nd International Conference on Software Quality, Reliability and Security Companion, QRS-C 2022
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 5 December 2022 through 9 December 2022
ER -