← All work

deep learning · federated learning · efficient adaptation

Federated Vision Transformers

A collaborative research codebase comparing centralized and federated Vision Transformer training on CIFAR-100 with DINO backbones, LoRA and TaLoS-style model editing.

Year
2025
Role
Collaborative implementation and experiment configuration

Collaboration note — Collaborative project. Repository hosted by a collaborator; contributions are presented as team work.

Federated learning diagram showing several non-IID clients fine-tuning Vision Transformer models and sending updates to a central aggregation server.
Diagram based on the repository architecture and reported evaluation path — not a fabricated product screenshot.

Problem

Federated visual learning combines non-IID data, distributed optimization and constrained client resources. The project explores whether parameter-efficient fine-tuning and model-editing strategies can be evaluated in a shared centralized/federated codebase.

What I built

The repository includes centralized and federated training scripts, a simulated client/server core, ViT models with DINO-pretrained backbones, LoRA and TaLoS-style editing modules, YAML configuration and Weights & Biases experiment tracking.

Research surfaceCentral + federated

The same codebase supports centralized and client-server experiments with multiple fine-tuning strategies.

Engineering decisions

Where the project becomes more than a demo.

The details that matter are usually in the constraints: who is allowed to decide, what gets measured, what is kept separate, and how failure is represented.

01

Keep centralized and federated paths comparable.

Shared model and configuration surfaces reduce the number of moving parts when comparing training modes.

02

Make adaptation strategy configurable.

Dense, LoRA and pruning/editing strategies can be switched through configuration rather than forked experiments.

03

Treat heterogeneity as part of the setup.

The client-server simulation is designed around non-IID client data and distributed aggregation instead of an IID toy assumption.

Evidence

What the repository actually supports.

3fine-tuning modes

dense, LoRA and TaLoS-style editing

ViTbackbone

DINO-pretrained Vision Transformer features

W&Btracking

sweeps and experiment logging integrated into the workflow

Limits

What I am not claiming.

  • This is collaborative academic research, not a production federated-learning platform.
  • The repository README does not support a single headline performance number, so the site does not invent one.
  • The work is presented with explicit collaborative attribution.

Signal

What this project demonstrates.

federated learningVision TransformersDINOLoRAmodel editing / pruningdistributed experimentation

Want to inspect the implementation?

Open repository ↗Ask me about it