Laminator Distributed (CS 858 Project)

This repository contains small demos of a distributed training coordinator + worker system implemented for the CS 858 project.

There are two example stacks in this repo:

• baseline/ — one coordinator and one (or more) worker containers for data-parallel-training but without any attestation or assumptions. See baseline/README.md.
• data-parallel-training/ — oone coordinator and one (or more) worker containers performing data-parallel training with signed, hashed tasks and results, assuming it was run inside TEEs. See data-parallel-training/README.md.

High-level flow

• The coordinator prepares training tasks that include the model architecture, initial weights, a batch of data, and training configuration.
• Tasks are dispatched to workers; workers perform local computation (compute gradients and a one-step local update across the configured number of epochs or steps) and return results to the coordinator.
• (for the second setup) Results are signed and hashed by workers; the coordinator verifies signatures and hashes, aggregates gradients, updates the global model, and records a verification report with the overall hashes that were input to the coordinator(data/hash_report.txt).
• Verifier can verify using the verify.py script in data/verifier.py. (also see the data/README.md)

Using Docker

From the respective folder run:

docker compose up --build

Stop and remove the stack with:

docker compose down

Quick edits

• Change the example model: edit the model code or model_string in coordinator/coordinator.py (or the coordinator folder for the baseline demo), then rebuild the images and restart the stack.
• Increase the worker count: add worker service entries to the compose YAML and set the coordinator --num-workers argument to the total number of worker services (see both demo READMEs for copy-paste examples).

Security & notes

• This repository include pre-generated public and secret keys under keys/ for convenience.

• baseline/README.md - try the data-parallel training.

• data-parallel-training/README.md — generate-keys helper, testing failure modes, and verification report behavior.