Research Aide– DSL – He, Pellin – 4.23.26

Argonne National Laboratory•Lemont, IL

10h•Onsite

About The Position

Federated foundation models (FMs) are crucial for scaling large models using decentralized private data. However, current FedFMs pipelines face challenges balancing privacy, communication efficiency, and model scaling. Existing differentially private federated learning methods often protect local data by perturbing gradients or model updates over many rounds, leading to accumulated privacy loss, reduced utility, and poor compatibility with large-scale foundation-model adaptation. This proposal introduces Conformal-DB enabled Federated Foundation Models, a scalable framework for differentially private FedFM training. It utilizes geometry-aware, distributional client interfaces. Each client maps local data to a foundation-model representation manifold. For non-IID data, a density-aware mechanism acts as a client-side geometric calibration layer, inspired by Conformal-DP, which minimizes unnecessary perturbation under heterogeneous densities. The client then sends privatized distributional model artifacts to the global aggregator. A privacy-aware federated scaling law is derived, detailing how achievable loss relates to model size, client count, per-client compute, communication budget, privacy temperature, and data density heterogeneity. This framework enables parameter-efficient FedFM training and paves the way for scalable, private foundation models with decentralized, manifold-structured data under non-IID conditions.

Requirements

Currently enrolled in undergraduate or graduate studies at an accredited institution, OR graduated from an accredited institution within the past 3 months, OR actively enrolled in a graduate program at an accredited institution.
Must be 18 years or older at the time the appointment begins.
Must possess a cumulative GPA of 3.0 on a 4.0 scale.
Must complete a satisfactory background check.
The entirety of the appointment must be conducted within the United States.

Nice To Haves

Experience with differentially private federated learning methods.
Familiarity with Conformal-DP.
Understanding of foundation models and their adaptation.

Responsibilities

Map local data into a foundation-model representation manifold.
Handle non-IID data using a density-aware mechanism as a client-side geometric calibration layer.
Release privatized distributional model artifacts to the global aggregator.
Contribute to research on federated foundation models.