NOMAI: Scalable Federated Learning with Private Weights

summary

The gist

Detailed Research Summary of NOMAI: Scalable Federated Learning with Private Weights NOMAI (Scalable MPC-based FL system) is a novel framework designed to simultaneously achieve model privacy, data

In short

NOMAI introduces a scalable Multi-Party Computation (MPC) system for Federated Learning that simultaneously ensures model privacy, data privacy, and data integrity. It solves the challenge of scaling FL to millions of users by using a committee-based architecture combined with Packed Shamir Secret Sharing and ML-aware optimizations. This allows it to achieve high performance while maintaining strong security guarantees.

Key concepts

Committee-Based MPC: Instead of one massive protocol, NOMAI divides users into small, disjoint committees. Each committee handles a specific shard of data and secret-shares that data among its members. This parallel structure allows many independent computations to run simultaneously, significantly improving scalability beyond traditional protocols.
Packed Shamir Secret Sharing (PSS): PSS is a cryptographic technique used within committees to enable SIMD-style parallelism. It efficiently shares secrets in a way that allows multiple data points or operations to be processed together in one step, drastically reducing the computational overhead associated with complex sharing schemes.
ML-Aware Optimizations: These are new cryptographic primitives tailored specifically for machine learning tasks. They include efficient linear layer optimizations and secure exponentiation protocols designed to minimize per-user computation time. These specific tools reduce the overall overhead of training models like MNIST by making operations faster.
Data Integrity Enforcement: NOMAI ensures data integrity by validating that individual user gradients meet necessary constraints, such as norm bounds, before aggregation. It cleverly performs the computationally intensive gradient computation in a 'semi-honest' mode and reserves the more rigorous malicious security checks only for the final validation steps.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 20 episodes
Retrieval-Augmented Generation (RAG) — explained in 19 episodes
Federated Learning — explained in 17 episodes
Flow Matching — explained in 17 episodes
Quantum Fisher Information (QFI) — explained in 16 episodes

This episode discusses

The paper

NOMAI: Scalable Federated Learning with Private Weights · Read on arXiv

Yiyi Sun, Mushtari Sadia, Amrita Roy Chowdhury

University of Michigan

The twin goals of privacy-preserving federated learning (FL) -- harnessing distributed data while protecting the parties who contribute it -- demands three properties: model privacy, data privacy, and integrity. Achieving all three in a realistic deployment -- with a single coordinator and millions of users -- remains an open challenge that prior work has yet to fully resolve. We present NOMAI, the first scalable MPC-based FL system to provide all three guarantees simultaneously. NOMAI operates under an honest-majority assumption in the semi-honest model, with an extension to the malicious-with-abort setting. At its core, NOMAI combines a multi-committee MPC architecture with packed Shamir secret sharing, enabling SIMD-style secure computation that scales far beyond what monolithic n-party protocols can support. We further introduce ML-aware optimizations and new cryptographic primitives that substantially reduce per-user overhead. Our evaluation shows NOMAI is practically deployable at unprecedented scale: it reduces per-user computation to just <350 ms (semi-honest) and <700 ms (malicious) for per iteration of training MNIST, achieving three orders-of-magnitude improvements over prior work while seamlessly supporting populations of over 106 users. This demonstrates, for the first time, that MPC-based FL can meet the scale demands of real-world systems without compromising on privacy or integrity.

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: I'm Nadia.

Elias: And I'm Elias, here with Priya.

Nadia: Today's paper: "NOMAI: Scalable Federated Learning with Private Weights".

Elias: Detailed Research Summary of NOMAI: Scalable Federated Learning with Private Weights NOMAI (Scalable MPC-based FL system) is a novel framework designed to simultaneously achieve model privacy, data privacy,

Nadia: First, who's behind it and why it matters.

Paper summary: Elias: To recap, NOMAI is presented as the first scalable MPC-based FL system that provides model privacy, data privacy, and data integrity all at once.

Nadia: The paper argues that achieving those three things in a realistic deployment with a single coordinator and millions of users remains an open challenge that previous work hasn't fully resolved.

Elias: They claim NOMAI solves this by combining a multi-committee MPC architecture with packed Shamir secret sharing to enable SIMD style secure computation.

Nadia: Specifically, they introduce ML-aware optimizations and new cryptographic primitives intended to substantially reduce the per user overhead associated with running this kind of protocol.

Priya: So for those of us focused on privacy and measurement, what is the main takeaway from this summary? What exactly does it claim about protecting our data?

Nadia: The claims are that for every honest user, no other user can learn any information about their raw private data while the central server only learns nothing beyond the final aggregate gradient.

Elias: They also claim model parameters held by the central server remain hidden from all participating users, which is weight privacy.

Priya: And integrity? How does this system guarantee that the gradients being sent are actually valid?

Nadia: Integrity is guaranteed by enforcing a validation predicate, Valid(cdot), which defines necessary constraints like norm bounds, and they claim this is done efficiently by executing the computationally expensive gradient computation under the cheaper semi-honest model.

Conclusion: Nadia: So, looking at the NOMAI paper by Yiyi Sun, Mushtari Sadia, and Amrita Roy Chowdhury, it's a system that tackles the scale problem directly in federated learning.

Elias: It moves beyond just building secure aggregation protocols to actually creating a system that can handle millions of users while keeping things private and honest.

Nadia: The implications are big because it shows MPC-based FL can be practically deployed at unprecedented scale without sacrificing the required security guarantees for data privacy and integrity.

Elias: It means we can start thinking about real world ML applications where large populations use federated learning, not just small test sets.

Priya: So, in simple terms for someone who only listens to this show, it means a tool exists that lets you train models collaboratively with millions of people while keeping your raw data and the model weights secret and accurate.

Nadia: Exactly. It's about moving beyond theoretical proof-of-concept to something that actually scales up in practice.

More episodes

← Home