A groundbreaking lightweight AI model has been developed to enable high-quality image generation without sending sensitive data directly to servers.
On Wednesdasy, the Ulsan National Institute of Science and Technology (UNIST) announced that a research team led by Yoo Jaejun of the Graduate School of Artificial Intelligence has developed PRISM (PRivacy-preserving Improved Stochastic Masking), an innovative federated learning AI model.
Federated learning is a cutting-edge technology that allows individual devices to train local AI models using sensitive data without uploading it to central servers. The results from these local models are then aggregated to create a single, powerful global AI.
PRISM functions as an AI intermediary, bridging the gap between local and global AI models during the federated learning process.
This revolutionary model boasts a 38% reduction in communication costs compared to its predecessors and operates at an ultra-lightweight 1-bit level, making it 48% smaller. This compact design allows PRISM to run on small devices like smartphones and tablets without taxing their CPU or memory resources.
Moreover, PRISM excels in scenarios where local AI models have significant disparities in data quality and performance. It accurately determines which local AI’s information is most reliable, resulting in high-quality final outputs.
A prime example of PRISM’s capabilities is its ability to transform a selfie into a Studio Ghibli-style image entirely within a smartphone. This eliminates the need to upload photos to external servers, addressing privacy concerns and delivering rapid results while safeguarding personal information.
While PRISM shows great promise, it’s worth noting that developing local AI models capable of generating images directly on smartphones remains a separate challenge. Experiments using standard AI benchmarks like MNIST, FMNIST, CelebA, and CIFAR10 demonstrated PRISM’s superiority in both reducing communication volume and improving image generation quality compared to traditional methods.
Further testing with the MNIST dataset confirmed PRISM’s compatibility with diffusion models, which are widely used in creating Studio Ghibli-style imagery.
The research team enhanced communication efficiency by implementing a binary mask method, selectively sharing only critical information instead of using a large parameter approach that shares all data indiscriminately.
To address data variance and learning instability, the team employed a sophisticated loss function MMD for precise evaluation of generation quality, coupled with a strategy MADA that intelligently aggregates contributions from each local AI.
Professor Yoo emphasized the broad applicability of this technology, stating that PRISM’s potential extends beyond image generation to text creation, data simulation, and automated documentation. It will become an effective and secure solution in fields handling sensitive information, such as healthcare and finance.
This groundbreaking research was a collaborative effort, with Professor Han Dong-Jun from Yonsei University contributing and UNIST researcher Seo Kyung-guk serving as the lead author.
The team’s findings have been accepted for presentation at ICLR (The International Conference on Learning Representations) 2025, one of the world’s top three AI conferences. The event is scheduled to take place in Singapore from April 24 to 28.
This innovative research was made possible through support from the South Korean Ministry of Science and ICT, the National Research Foundation of Korea, the Institute of Information and Communication Technology Planning and Evaluation, and the UNIST Supercomputing Center.