SoftBank Deploys Nvidia Tech for Telecom AI Data Platform

San Jose, CA / Tokyo, Japan – SoftBank Corp has launched a synthetic data generation platform using Nvidia Corp's technology to train its Large Telecom Model, a generative AI foundation model developed for the telecommunications sector, the company said Tuesday.

The platform employs the Nvidia NeMo Safe Synthesizer and incorporates differential privacy techniques, enabling the use of network operation and base station configuration data for AI training while protecting confidential information.

SoftBank said previous use of such network data was constrained by security concerns.

Traditional anonymization methods, the company noted, reduced the data's utility for AI training by removing the complex correlations needed to detect system faults and network quality issues.

According to the company, the system is designed to ensure that adding or removing individual data points does not significantly affect the AI model’s output. 

The generated datasets are tested for resistance to membership and attribute-inference attacks to prevent the extraction of sensitive information, including base-station locations.

SoftBank reported that tests of the synthetic data confirmed its accuracy for assessing network quality trends and operational data. 

The company plans to conduct further evaluation with telecom operators, network equipment manufacturers, and academic institutions. It also intends to introduce safeguards for large language models in AI inference and contribute the technology to industry consortia such as the AI-RAN Alliance.