FedOps 1.2 Announcement
What is FedOps ?
Problem Statement: Organizations need Federated learning (FL) to learn from sensitive, siloed, or geo-restricted data (health, finance, IoT) without moving it—meeting privacy laws (e.g., HIPAA/GDPR), data-residency rules, and central storage risk. It also enables better, more robust models by utilizing diverse non-independent and identically distributed (non-IID) data across organizations/sites, supports on-device personalization, and provides continuous collaboration where raw-data sharing is impossible.
Solution: FedOps is a lifecycle management platform for Federated Learning (FL) that integrates the Flower framework. It helps teams/organizations train, monitor, and evaluate AI models collaboratively across heterogeneous computing nodes among organizations—safely and efficiently—without centralizing sensitive raw data among others.
Now with the stable release of FedOps 1.2, managing and scaling privacy-preserving AI became more easier and powerful.

What is Flower?
Flower is an open-source framework for Federated Learning (FL), introduced by Flower Labs from University of Cambridge. It provides the core client–server protocol, strategies (e.g., FedAvg and variants), and tooling to run FL across many devices or silos—so you can train models without moving raw data off each client.
So, what does FedOps bring over Flower here?
FedOps is a lifecycle operations platform built on top of Flower. In simple analogy, if Flower is the FL “engine,” FedOps is the vehicle + dashboard + pit crew:
- Orchestrate Flower jobs end-to-end (config → launch → monitor → evaluate)
- Automate boilerplate (generate validated configs and runnable FL server/client stubs)
- Operate at scale with observability, dashboards, logs, client selection, XAI, Clustering & Hyperparameter optimization (HPO), LLM training, and multimodal aggregation.
In short: Flower gives you the FL protocol and primitives; FedOps turns them into a production workflow with automation, visibility, and extras for non-IID, interpretability, LLM, and multimodal scenarios.
🧱 What was in FedOps 1.1 ?
Foundation modules that made FL practical at scale:
- FL Scalize
- Facilitates the integration and deployment of data and models in FL environments.
- Supports frameworks such as Flower by simplifying the client/server architecture setup for FL applications.
- Manager
- Oversees the real-time operation and coordination of federated learning sessions.
- Manages both client and server-side processes, ensuring synchronized execution and lifecycle control.
- Contribution Evaluation & Client Selection (CE/CS)
- Evaluates the contribution of each FL client based on performance metrics.
- Dynamically selects optimal clients to participate in training rounds, improving efficiency and fairness.
- Contribution Evaluation & Client Selection (CE/CS)
- CI/CD/CFL Integration
- Enables automated code deployment and scheduling by linking to version control repositories (e.g., Git).
- Supports continuous integration and delivery tailored to the needs of federated training cycles.
- Monitoring & Dashboard
- Provides visual dashboards for real-time monitoring of FL system status.
- Displays training progress, model updates, client participation, and system performance across the lifecycle.
What’s new in 1.2?
- Adapt LLMs to distributed data without sharing raw data—fully privacy-preserving. Powered by FlowerTune, FedOps 1.2 automates the end-to-end pipeline: task config → distributed LoRA training → model aggregation → global checkpointing.
- Solves: High GPU/comms overhead, data silos, manual orchestration.
- Delivers: Parameter-efficient tuning, minimal setup, domain-specific LLM adaptation at scale.
✨ Enhanced Automatic Configuration and FL Server Code Generation
- FedOps 1.2 auto-generates validated configs + runnable FL server/client code stubs instantly from your task spec. No more manual tuning—strategy, metrics, hooks, and datasets are pre-wired. Deployment just got drastically faster.
🔬 Advanced Federated Learning Capabilities – Now Fully Turnkey via Simple Config Flags
- Explainable AI (XAI) Built-In Enable federated interpretability with Grad-CAM-based XAI to visualize model decisions on local physiological or image data. Clients generate Grad-CAM heatmaps (e.g., MNIST), report aggregated metrics (entropy, similarity), plus per-round explainability hooks for feature importance, drift checks, and exportable reports. Close the interpretability gap in privacy-preserving FL.
- Intelligent Client Clustering + Hyperparameter Optimization (HPO) Automatically group clients by data/behavioral signatures, then run cluster-specific HPO to unlock optimal performance even under severe Non-IID conditions.
- FedMAP Aggregation for Multimodal FL (MMFL) A new multimodal FL aggregation method that dynamically learns adaptive client weights from interpretable meta-features—engineered for real-world multimodal, non-IID clients.
All features ship with full guided tutorials, end-to-end examples, and production-ready use cases.
- We’ve integrated a real-world federated IoT health pipeline using Fitbit wearable data. Enables privacy-preserving sleep-quality prediction and personalized health monitoring without centralizing user data.
- Introduces an open-source lightweight SleepLSTM model with 3-layer LSTM + projection bottleneck for temporal feature learning on multivariate Fitbit signals.
🖥️Enhanced FL Server Logs & Monitoring
- Problem Solved: Hard to track system health, performance drift, and errors in production FL runs.
- What’s New: Deep observability with real-time insights into metrics, logs, and lifecycle state.
- FL Status Tracking: Monitors creation → execution → termination stages, providing clear status indicators to guide next actions.
- Live Server Operation Logs: Stream server logs directly to the web dashboard—view errors, training progress, and learning status as it happens.
👉 No more blind runs. Full visibility, instant debugging.
Installation Procedure & Requirements
**pip install fedops**
This release is a valuable advancement for researchers, engineers, and teams building privacy-first AI on distributed data—delivering production-grade orchestration, LLM adaptation, multimodal robustness, and full observability with minimal-touch deployment.
Thanks to our Contributors
We extend our sincere gratitude to the dedicated team at Gachon Cognitive Computing Lab who made this release possible:
- Yong-gyom Kim (gyom1204@gachon.ac.kr)
- InSeo Song (z8086486@gachon.ac.kr)
- JingYao Shi (11490800573@gachon.ac.kr)
- Ri-Ra Kang (rirakang@gachon.ac.kr)
- Akeel Ahamed(akeelahamed@gachon.ac.kr)
- MinHyuk Jung(bvnm0121@gachon.ac.kr)
- JinYong Jung (wlsdyd5373@gachon.ac.kr)
- Advised by Prof. KangYoon Lee (keylee@gachon.ac.kr)
Explore the new features and documentation on our website
(link: https://ccl.gachon.ac.kr/fedops).
Join the Discussion: Connect with the community on our Slack channel