Projects

Project Details

Blockchain-Based Federated Learning with Adaptive Differential Privacy 2025

Project Overview

This project implements a Blockchain-based Federated Learning (BC-FL) system with Adaptive Differential Privacy that eliminates the need for a trusted central aggregator. The system combines Ethereum smart contracts for decentralized coordination, IPFS for distributed model storage, and adaptive differential privacy to optimize the privacy-utility trade-off.

The key innovation is replacing the traditional central FL server with a blockchain-based coordination protocol, while implementing an adaptive noise mechanism that dynamically adjusts privacy parameters based on training progress, optimizing both model utility and privacy budget consumption.


System Architecture

Three-Layer Architecture

  • Compute Layer (Client-Side):

    Local model training with PyTorch, adaptive noise injection using Opacus, gradient clipping and privacy accounting, IPFS upload/download operations
  • Storage Layer (Off-Chain):

    IPFS for decentralized model storage, content-addressed storage (CID-based), tamper-proof model versioning
  • Coordination Layer (On-Chain):

    FLRegistry.sol smart contract (Ethereum/Ganache), round management and state tracking, privacy budget enforcement, client reputation system

Key Features

Privacy Mechanisms

  • Centralized DP (CDP)
  • Local DP (LDP)
  • Adaptive DP
  • Adaptive Clipping

Blockchain Features

  • Client registration
  • Reputation tracking
  • Privacy budget enforcement
  • IPFS hash verification

Technical Details

Model Architecture

Dataset: CIFAR-10 (32×32 RGB images, 10 classes). Lightweight CNN with 2 convolutional layers, 2 fully connected layers, and GroupNorm for DP compliance.

Privacy Accounting

Custom Rényi Differential Privacy (RDP) accountant that tracks privacy budget across multiple rounds with varying noise levels, converting RDP to (ε, δ)-DP using optimal alpha selection.

Adaptive Mechanism

Monitors validation loss after each epoch and dynamically adjusts noise multiplier based on loss trend. Loss improving → reduce noise (spend budget), Loss stagnating → increase noise (conserve budget).


Project Information

Course

EECE 571B - Blockchain Foundations

Dataset

CIFAR-10

Stack

Python, PyTorch, Solidity, Ethereum, IPFS, Opacus, Foundry, Ganache

Technologies

Federated Learning, Differential Privacy, Blockchain, Smart Contracts, Distributed Storage

Links
GitHub Repository

Research Paper

Key Results

Privacy-Utility Trade-off

Adaptive DP achieves better accuracy for similar privacy budgets compared to static approaches.

Decentralization Cost

Local DP (decentralized) has lower utility than Centralized DP but eliminates trust requirements.

Economic Feasibility

Mainnet costs are prohibitive (~$17/round), but Layer 2 solutions reduce costs by 95%.