Federated Learning for Privacy-Preserving Model Training

A decentralized training simulation with secure aggregation concepts and cryptographic audit logging.

This project demonstrates how distributed clients can collaboratively train a machine learning model without sharing raw data, using the Flower (FLWR) Federated Learning framework. It uses the UCI HAR dataset split across five clients and introduces SHA-256 audit logging for tamper-evident update tracking. Accuracy improves consistently across all federated rounds.

---

📌 Features

• Federated Learning with FedAvg
• Conceptual secure aggregation for privacy
• SHA-256 audit logging to verify model update integrity
• Uses UCI HAR dataset (561 features, 6 activity classes)
• Five-client decentralized training setup
• Accuracy improves from 78.11% → 92.76% over 5 rounds

---

📂 Dataset

The model is trained on the UCI Human Activity Recognition (HAR) dataset, which contains accelerometer and gyroscope signals from smartphones. Each sample has 561 features and one of six activity labels. The training set is partitioned into five subsets, simulating federated IoT clients.

---

🧠 Model Architecture

Each client trains a lightweight neural network implemented in PyTorch:

• Input: 561-dimensional feature vector
• Hidden layer: 100 neurons + ReLU
• Output: 6-class softmax

This architecture is defined in the Net model.

---

🔄 Federated Learning Workflow

From the project methodology:

1. Server sends global model to all clients
2. Clients train locally for one epoch
3. Clients send updated weights
4. Server aggregates using FedAvg
5. Repeat for 5 rounds

---

📊 Results

Results from the evaluation table:

Round	Accuracy (%)	Improvement
1	78.11	–
2	87.22	+9.11
3	90.03	+2.81
4	91.62	+1.59
5	92.76	+1.14

The results show steady accuracy improvement as global updates aggregate client contributions.

---

🔐 Audit Logging

The project includes SHA-256 hashing of model weights after each round, creating a tamper-evident audit trail. This provides blockchain-style integrity without requiring an actual blockchain implementation. (See audit.py for details.)

---

▶️ How to Run

1. Prepare the dataset

python prepare_data.py

2. Start the federated server

python server.py

3. Launch each client (open 5 terminals)

python client.py 1
python client.py 2
python client.py 3
python client.py 4
python client.py 5

---

Future Improvements

• Real blockchain integration for on-chain audit
• Implement secure aggregation (AES/SMPC)
• Support Non-IID data distributions
• More complex neural models

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

If you want a shorter GitHub description, a banner, or badges, I can generate those too.