GraphMind

Distributed GNN training with Byzantine fault tolerant consensus.

Overview

GraphMind implements distributed graph neural network training with three core components:

1. TA-BFT Consensus - Topology-aware Byzantine fault tolerant consensus using eigenvector centrality for validator weighting
2. Adaptive Graph Partitioning - Multi-level partitioning with spectral clustering and simulated annealing
3. FedTopo Aggregation - Federated learning with graph Laplacian regularization

Architecture

Edge Nodes (Local GNN Training)
            │
            ▼
    TA-BFT Consensus Network
            │
            ▼
    Global Model Aggregation
    ├── AMP Partitioner
    ├── FedTopo Aggregator
    └── PORA Optimizer

Installation

git clone https://github.com/elcruzo/GraphMind
cd GraphMind

conda create -n graphmind python=3.9
conda activate graphmind

pip install -r requirements.txt

# MPI for distributed training
sudo apt-get install libopenmpi-dev
pip install mpi4py

Usage

Distributed Training

# Single machine
mpirun -np 8 python src/distributed_train.py --config config/training_config.yaml

# Multi-machine cluster
mpirun -f hostfile -np 32 python src/distributed_train.py --config config/training_config.yaml

Byzantine Simulation

python src/byzantine_simulation.py --failures 2 --nodes 10 --graph cora

Production Mode

python distributed_node.py --config config/node_config.yaml --node-id node1

Configuration

aggregator:
  method: "topology_aware"  # topology_aware, coordinate_median, trimmed_mean
  regularization_strength: 0.1
  personalization_rate: 0.3
  byzantine_robust:
    trim_ratio: 0.2

Algorithms

TA-BFT Consensus

• Incorporates graph topology into Byzantine agreement
• Eigenvector centrality weighting for validator selection
• Maintains safety under f < n/3 Byzantine failures
• O(n^2) message complexity, O(log n) round complexity

Adaptive Multi-Level Partitioning

• Dynamic repartitioning based on community evolution
• Multi-objective optimization: cut-size, load balance, communication cost
• Spectral clustering + simulated annealing

FedTopo

• Structure-aware parameter aggregation using graph Laplacian
• Personalized GNN models with global consensus
• Convergence guarantees under non-IID data

Performance

Dataset	Nodes	Byzantine	TA-BFT Rounds	PBFT Rounds	Speedup
Cora	100	10%	3.2	8.1	2.5x
CiteSeer	500	20%	4.7	12.3	2.6x
PubMed	1000	30%	6.1	18.7	3.1x

Requirements

• Python 3.9+
• PyTorch 2.0+
• torch-geometric 2.4+
• NetworkX 3.0+
• OpenMPI (for distributed training)
• CUDA 11.8+ (optional, for GPU)

License

MIT