Chester Chen | NVIDIA, Senior Product and Engineering Manager

Holger Roth | NVIDIA, Principal Federated Learning Scientist

Peter Cnudde | NVIDIA, Director of Engineering

Chester Chen | NVIDIA, Senior Product and Engineering Manager

Ziyue Xu | NVIDIA, Senior Data Scientist

Chester Chen | NVIDIA, Senior Product and Engineering Manager

In the LLM-agent era, federated learning must evolve from aggregating numerical gradients to collaborating through language. To explore this shift, Xiaoxiao Li (UBC/Vector Institute) introduces two new paradigms: FedTextGrad, a collaborative prompt optimization framework that aggregates LLM-generated textual feedback (while exposing new safety vulnerabilities to malicious instruction injection), and MemCo, a collaborative memory framework that allows agents to share transferable global workflows without contaminating local environments. Together, these approaches redefine federated collaboration by replacing gradient exchange with language-based sharing, raising critical new questions about capability and trust. 

Xiaoxiao Li | The University of British Columbia; Vector Institute, Associate Professor, Electrical and Computer Engineering; Canada CIFAR AI Chair

BAT-HI—Better AI Together for Health Innovation—is an ABAC, APEC Business Advisory Council, project designed to advance trusted, inclusive, and collaborative healthcare AI through federated learning. As health systems across the APEC region face growing demands, fragmented data, privacy concerns, and uneven AI capacity, BAT-HI proposes a new model of cooperation: moving from data sharing to intelligence sharing. By leveraging Quanta QOCA’s AI medicine platform and connected health ecosystem together with NVIDIA FLARE’s federated learning framework, BAT-HI enables AI models to learn across distributed hospitals, research institutions, and economies while allowing sensitive health data to remain local. This approach supports privacy protection, institutional trust, and cross-border collaboration without requiring centralized data sharing.

Ted Chang | Quanta Computer, Taiwan Representative to ABAC, CTO

Health biobanks around the world hold extraordinary potential for population-scale discovery, yet privacy regulation, institutional data-governance frameworks, and national data-sovereignty requirements make physically pooling these cohorts into a single dataset legally and practically impossible. Trusted Research Environments (TREs) resolve part of this tension by holding sensitive data behind tightly controlled ingress and egress—but that same security posture makes cross-cohort, cross-border collaboration genuinely hard. Federated analysis offers a path forward: the computation travels to the data, and only privacy-preserving aggregate results ever leave each environment.

Anders Dale | J. Craig Venter Institute, Professor

The Infocomm Media Development Authority (IMDA) Singapore has been supporting businesses in deriving valuable insights from data while protecting privacy and safeguarding commercially sensitive information through practical experimentation with privacy-enhancing technologies. This session will share an overview of the IMDA Privacy-Enhancing Technology Sandbox, its offerings, and use cases. As part of its efforts to promote adoption, IMDA also publishes technical guides, and this session will introduce the recently published “Guide to Federated Learning,” which helps organizations assess the suitability of federated learning for their needs and understand key implementation considerations.

Khalid Ahmad | Personal Data Protection Commission (PDPC) Singapore, Acting Deputy Director

Zhongyuan Huang | Infocomm Media Development Authority of Singapore (IMDA), Senior Manager

Mobile base stations are evolving into distributed platforms for edge AI, where 6G Integrated Sensing and Communication (ISAC) allows towers to extract range-Doppler sensing data directly from wireless signals. However, varying geometry, traffic, and propagation conditions can cause local anomaly detection models to fail. In this session, Dr. Mauro Belgiovine demonstrates how federated learning and NVIDIA FLARE enable base stations to collaboratively train robust anomaly detectors without sharing raw data. Using multi-tower simulations powered by NVIDIA Sionna™ ray tracing (SionnaRT), the talk explores the critical tradeoffs between local specialization, global generalization, personalization, and cross-domain robustness.

Mauro Belgiovine | NVIDIA, Senior Software Engineer

Unified multimodal models (UMMs) are emerging as strong foundation models that can do both generation and understanding tasks in a single architecture. However, they are typically trained in centralized settings where all training and downstream datasets are gathered in a central server, limiting the deployment in privacy-sensitive and geographically distributed scenarios. In this paper, we present FedUMM, a general federated learning framework for UMMs under non-IID multimodal data with low communication cost. Built on NVIDIA FLARE, FedUMM instantiates federation for a BLIP3o backbone via parameter-efficient fine-tuning: clients train lightweight LoRA adapters while freezing the foundation models, and the server aggregates only adapter updates. We evaluate on VQA v2 and the GenEval compositional generation benchmarks under Dirichlet-controlled heterogeneity with up to 16 clients. Results show slight degradation as client count and heterogeneity increase, while remaining competitive with centralized training. We further analyze computation--communication trade-offs and demonstrate that adapter-only federation reduces per-round communication by over an order of magnitude compared to full fine-tuning, enabling practical federated UMM training. This work provides empirical experience for future research on privacy-preserving federated unified multimodal models.

Jindong Wang | William & Mary, Assistant Professor

Federated learning enables multiple organizations to collaboratively train a model without sharing their private data. However, when the model is a large language model (LLM), standard federated learning becomes impractical for resource-constrained participants due to the high computational cost of LLM training or fine-tuning. Federated Split Learning (FSL) addresses this challenge by partitioning the model between clients and a server, allowing clients to execute only a small portion of the model locally. However, applying FSL to LLMs introduces a fundamental privacy paradox: the autoregressive nature of LLMs makes input leakage through transmitted activations unavoidable, while existing perturbation-based defenses are largely ineffective. In this talk, we propose a novel framework for privacy-preserving federated split learning of LLMs. Experimental results show that our approach provides strong privacy protection with only modest utility loss and system overhead, making FSL-based LLM fine-tuning a practical solution for collaborative LLM training.

Wenjing Lou | Virginia Tech, Professor

Foundation models are enabling computational oncology models to be adapted to local clinical tasks with institution-specific labeled data, but rare cancers, including many pediatric cancers, expose a critical bottleneck: no single center has enough cases to reliably tune or validate models. This challenge is especially urgent for pediatric brain tumors, the leading cause of cancer-related mortality in children. We developed a federated framework that leverages pathology foundation models to predict recurrence risk in multi-institutional pediatric brain tumor cohorts across the United States, Europe, and Asia, enabling accurate risk stratification to guide surveillance intensity, treatment escalation, and follow-up planning. By allowing sites to contribute to shared model learning while retaining patient data locally, this work provides a scalable path toward clinically actionable AI for rare diseases.

Junhan Zhao | University of Chicago, Assistant Professor

Federated learning on tabular data faces a question model aggregation can't answer: which features matter, when no site can see another's data columns? Feature Election, now contributed to NVIDIA FLARE, solves this in three phases. Each client runs a local selector (lasso, elastic-net, random forest, mutual information and more) and shares only feature votes and scores — never raw data. The server aggregates votes by weighted voting and elects a global feature mask which will be used by every participant, then coordinates FedAvg training on the reduced set. The freedom degree hyper parameter decides which features are selected, from the union, up until the intersection of selected features and it can be tuned using an iterative hill climbing method. Originating from FLASH (Best Student Paper, IEEE FLTA 2025), this talk covers the design, freedom degree optimization, the intersection–union trade-off, and how Feature Election became a production-ready contributi

Ioannis Christofilogiannis | Trinity College Dublin, MSc Intelligent Systems

Federated learning lets healthcare institutions build AI models without centralizing patient data—but successful distributed training is only the beginning of clinical translation. To generate real healthcare impact, federated models must be externally validated, embedded in clinical workflows, linked to predefined clinical actions, evaluated prospectively, and monitored throughout deployment. Chin Lin will introduce a clinical AI learning cycle connecting distributed signal discovery with pragmatic trials and real-world implementation. Drawing on our experience with multinational federated learning across AI-enabled electrocardiography, chest radiography, and electronic health records, and on more than ten pragmatic AI randomized trials, this session will show how routinely collected medical data can surface latent disease signals across institutions while local data governance is preserved—and why predictive accuracy alone does not establish clinical utility. We will define an indication–action pathway that converts an AI prediction into measurable change in diagnosis, treatment, and patient outcomes, and show how pragmatic trials and lifecycle monitoring fit into a federated healthcare AI program. We will close with deployment across hospitals, emergency medical services, public health centers, and home-care environments, and how platforms such as NVIDIA FLARE support not only privacy-preserving model development but cross-site validation, lifecycle learning, and sustainable clinical AI implementation

Chin Lin | National Defense Medical University, Professor

Federated learning enables multiple parties to collaboratively train machine learning models without sharing raw data, but standard implementations remain vulnerable to privacy leakage through model updates, gradients, and aggregation steps. This presentation explores how secure multi-party computation (MPC) can be integrated into federated learning pipelines to provide stronger, cryptographically grounded privacy guarantees. We examine how MPC protocols allow participants to jointly compute model updates without exposing individual contributions, even to the central server or coordinating party. The presentation will delve into trade-offs between privacy, computational overhead, and communication costs.

Victor Sucasas | Technology Innovation Institute (TII), Executive Director, Cryptography engineering

Chester Chen | NVIDIA, Senior Product and Engineering Manager