SoftBank Corp. Develops Secure Access Feature
to Edge AI Servers on “AITRAS” Utilizing Local
Breakout Technology

(1) LLM switching application utilizing URSP

To handle highly confidential corporate information with LLMs, companies need to build their own proprietary LLMs. By deploying these proprietary LLMs within the edge AI servers on AITRAS, organizations can securely manage sensitive data. At the same time, various LLMs are emerging rapidly, with continuous advancements in performance. By effectively switching between proprietary LLMs and external LLMs, businesses can enhance operational efficiency and improve usability.

SoftBank developed an application that enables access to both public and private LLMs by utilizing URSP to simultaneously configure PDU sessions for both the Internet and the edge AI servers on AITRAS. Furthermore, SoftBank implemented an AI agent that operates on the device using a lightweight LLM to assess the confidentiality of user input. This allows for an automatic switching mechanism between public and private LLMs without requiring users to manually select which LLM to use. With this approach, employees using company-issued work devices can leverage LLMs while ensuring compliance with corporate security policies, regardless of individual awareness of confidentiality requirements.

(2) Multimodal RAG application utilizing LADN

As AI adoption expands across industries, its potential applications in manufacturing are also evolving. For example, AI could autonomously understand the overall factory environment, issue instructions to production equipment, and verify information without human intervention. To support such use cases in the manufacturing sector, SoftBank developed an LLM application compatible with LADN. This application enables PDU sessions only within designated areas, such as factory premises, allowing access to the edge AI server on AITRAS exclusively within the factory. By restricting LLM access to on-site use, this solution enhances security and ensures a highly controlled AI deployment environment.

Additionally, SoftBank built a multimodal RAG (Retrieval-Augmented Generation) system on the edge AI server of AITRAS, which is capable of processing various types of data, including sensor readings from factory equipment and surveillance camera footage.
By leveraging multimodal information such as video, audio, and sensor data, this system enables more precise responses to user inquiries, enhancing the accuracy and reliability of AI-driven decision-making in industrial environments.

Moving forward, SoftBank will continue to develop new features and expand use cases by integrating AI-RAN with mobile network technologies, contributing to solving social challenges through AITRAS.
Additionally, SoftBank published a white paper that includes details on this technology. For more information, please refer to the following link (https://www.softbank.jp/en/corp/technology/research/news/065/).

Hideyuki Tsukuda, Executive Vice President & CTO at SoftBank Corp. said: “This demonstration of concrete solutions through the integration of AI and mobile network technology has highlighted the critical role of mobile networks in an AI-integrated society. To serve as the foundation for an AI-integrated society, it is crucial to focus not only on the implementation of 'AITRAS' as a Radio Access Network (RAN) but also on end-to-end design. Moving forward, we will drive innovation centered around 'AITRAS' while considering end-to-end solutions, aiming to develop high-value social infrastructure.”