[3B3-GS-11] AI and Society: reliability

[3B3-GS-11-04] A study on the Identification of AI

〇Ryohei Suzuki¹, Nami Ashizawa¹, Naoto Kiribuchi¹ (1. NTT Social Informatics Laboratories)

Keywords:AI, AI Security, Authentication

The authenticity of software is validated by its identity. For example, a user can use a hash value of the software as its identity. On the other hand, when we want to validate the authenticity of AI, it is not obvious whether the conventional concept of software authenticity validation is appropriate. This is because some neural network-based AIs continuously learn and update their internal parameters to change their output, and the changes in these AI models are not always what the creators intended.
However, these kinds of changes in AI can also be considered as autonomous evolution of AI, and if it does not contradict the intention of its creator, it will have the same identity as before the changes. In this paper, we propose levels of AI identity assurance with reference to existing authentication frameworks, and consider what level of assurance is required in multiple use cases.