The Need

Federated learning enables collaborative AI training without sharing sensitive data, making it attractive for healthcare, finance, edge computing, and enterprise applications. However, widespread deployment remains limited by the enormous communication burden associated with repeatedly transferring large model updates between distributed clients and a central server. As modern AI models scale to billions of parameters, communication costs can become the dominant bottleneck, increasing latency, infrastructure requirements, energy consumption, and overall deployment costs. Existing communication-reduction methods still remain fundamentally dependent on model size.

The Technology

OSU engineers and their collaborators have developed a novel federated learning framework that dramatically reduces communication requirements by replacing the exchange of full model updates with a compact representation based on zeroth-order optimization. Rather than transmitting high-dimensional parameter vectors, clients and servers exchange only a small amount of information that enables reconstruction of the learning process while preserving model performance. The approach decouples communication cost from model size, allowing efficient distributed training and fine-tuning of very large AI models. Experimental results demonstrate substantial communication savings while maintaining competitive learning accuracy across a range of machine learning tasks.

Commercial Applications

• Federated training and fine-tuning of large language models (LLMs) across distributed organizations.
• Privacy-preserving AI systems for healthcare, financial services, and regulated industries.
• Edge AI and IoT deployments with limited network bandwidth.
• Distributed machine learning platforms for cloud service providers, AI infrastructure vendors, and enterprise AI developers.

Benefits/Advantages

• Dimension-free communication
• Substantially reduced bandwidth consumption
• Maintains competitive model performance
• Supports privacy-preserving distributed AI