Project Octopus - Quantum fingerprinting for securing generative models

点击翻译按钮获取中文摘要

AI models have rapidly become the basis of an organization's core operation and repository of its intellectual property. In the case of an enterprise, models can be their very business model. In addition, with the costs of training generative AI systems potentially running into the millions or 10's of millions of dollars, there has been increasing interest in developing methods to prevent the extraction or theft of a trained model. Even when a model is hosted on a device controlled by the model's creator, it may be vulnerable to hackers through gaps in the model creator's security systems, and there is no technical protection against accidental leaks or malicious acts by employees. Companies must resort to legal protections only to protect their models from these leaks.We propose a quantum fingerprinting solution for securing trained generative models from hackers or from accidental or malicious leaks.This project will be led by ORCA Computing, who were formed in October 2019 as a spin-off from Professor Ian Walmsley's ultra-fast quantum optics group at the University of Oxford and is now based in West London with over 40 employees.