P2P-Fed-Learning

P2P Federated Learning is a decentralized platform that connects ML users who need affordable training with trainers who have idle compute power (CPUs/GPUs). Instead of relying on expensive cloud providers, researchers can upload their models and datasets, which are then split, distributed, and trained across a peer-to-peer network of independent nodes.

The system combines:

Hedera blockchain for transparent payments, audit logs, and smart contract–based task management.

Akave O3 decentralized storage with presigned URLs for efficient and secure dataset/model distribution.

py-libp2p networking to enable peer-to-peer coordination and communication between trainer nodes.

Why it matters

For ML users: Train models without expensive infrastructure—just connect a Hedera wallet and start.

For trainers: Monetize idle gaming rigs or compute resources by competing to train model chunks, creating a market for cost-efficient ML.

We built P2P Federated Learning by stitching together decentralized tech and machine learning orchestration into a single system.

Frontend (React + Node.js) A clean interface for ML users to upload their datasets and models, monitor training progress, and receive the final weights. The frontend also handles environment setup with API keys and JWT tokens for secure access.

Decentralized Storage (Akave O3) Datasets are split into chunks and uploaded to Akave O3. Instead of moving massive files across the network, we generate presigned URLs that trainers can fetch directly. Trainers also upload trained weights back to Akave, again via presigned URLs, keeping transfers efficient and access controlled.

Blockchain Layer (Hedera Hashgraph) Hedera smart contracts manage training tasks, escrowed payments, and completion events. Consensus Service topics are used for real-time logging of trainer activity, so even if nodes fail locally, their logs are still accessible and verifiable on-chain.

P2P Networking (py-libp2p) Trainer nodes communicate over a libp2p-based mesh network. This enables decentralized job orchestration—ML tasks are published into the mesh, trainers compete to take them, and results flow back through the same network.

Training Flow

User uploads model + dataset.

Dataset chunks + presigned URLs are published into the network.

Trainers pick jobs, train locally, and upload weights back to Akave.

Hedera contracts verify state and handle payments.

The user retrieves encrypted weights URLs and integrates the trained model.

The result is a trust-minimized, low-cost federated learning platform where anyone can plug in as a trainer or ML user, without relying on centralized cloud providers.