A Study on Improving Trust of Model Cards

Abstract

MLaaS and the distribution of pre-trained models have reduced the difficulty of utilizing AI. This has made it convenient and easy to implement AI-embedded services. At the same time, there are more opportunities to use AI as a user. While the use of AI is becoming more accessible to many businesses, the issue is whether the AI provided is what the user intends, and the trust of AI use. As an example of ensuring the trust of AI, the use of model cards that clearly indicate supplementary information such as the developer of the AI and the training dataset has been advocated. In this paper, we focus on the fact that no method has validated the trust of the information described in model cards. Referring to the existing method in Web PKI as an existing method to establish information trustworthiness, the relationship between model cards and AI is analogous to the relationship between Web PKI certificates and Web servers, and the requirements necessary to ensure information trustworthiness are organized.