CSC8370 Data Security Project: README
Project Overview
This project implements machine learning solutions for MNIST handwritten digit classification. The project is divided into three levels, each with increasing complexity:
	•	Level 1: Centralized Learning
	•	Level 2: Federated Learning
	•	Level 3: Robust Federated Learning with Malicious Client Detection
The models are evaluated on accuracy, and output files are generated to facilitate further analysis.

Team Members
	•	Nagamedha Sakhamuri – 002828574 – nsakhamuri1@student.gsu.edu
	•	Pranav Chowdary Nagothu – 002852538 – pnagothu1@student.gsu.edu
Contributions: Both team members collaborated equally on all levels, including coding, testing, and documentation.

Project Structure
The project consists of three Python scripts:
	1	Level1_Final.ipynb: Implements centralized learning for MNIST classification.
	2	Level2_Final.ipynb: Implements federated learning with 10 clients.
	3	Level3_Final.ipynb: Implements federated learning with malicious client detection and robust aggregation.

Requirements
The project requires the following Python libraries:
	•	torch
	•	torchvision
	•	matplotlib
	•	random
	•	copy
	•	google.colab (for file downloads in Colab)
Install missing libraries using:
!pip install torch torchvision matplotlib

Instructions
Setup
	1	Open Google Colab.
	2	Upload the desired level script (Level1_Final.ipynb, Level2_Final.ipynb, or Level3_Final.ipynb).
	3	Run the script step by step or all at once.
	4	Ensure the runtime type is set to GPU for faster execution:
	◦	Navigate to Runtime > Change runtime type.
	◦	Set Hardware Accelerator to GPU.

Level 1: Centralized Learning
	•	Overview: All MNIST data is centralized, and a Convolutional Neural Network (CNN) is trained to classify the digits.
	•	Execution:
	1	Open Level1_Final.ipynb in Google Colab.
	2	Run the script to train the model.
	•	Output Files:
	◦	mnist_training.log: Logs training and testing accuracy for each epoch.
	◦	best_mnist_model.pth: Stores the best-performing model's parameters.
	◦	mnist_performance_graph.png: Visualizes accuracy over epochs.

Level 2: Federated Learning
	•	Overview: Implements federated learning with 10 clients, each holding a partition of the MNIST dataset. Models are trained locally and aggregated on a central server.
	•	Execution:
	1	Open Level1_Final.ipynb in Google Colab.
	2	Run the script to simulate federated learning.
	•	Output Files:
	◦	level2_training.log: Logs accuracy values for each global epoch.
	◦	federated_mnist_model.pth: Stores the aggregated global model's parameters.
	◦	performance_graph_level2.png: Visualizes accuracy over federated learning rounds.

Level 3: Robust Federated Learning with Malicious Client Detection
	•	Overview: Introduces malicious clients that submit falsified updates. Implements mechanisms to detect and mitigate the effects of such clients using robust aggregation.
	•	Execution:
	1	Open Level1_Final.ipynb in Google Colab.
	2	Run the script to simulate federated learning with malicious client detection.
	•	Output Files:
	◦	federated_mnist_level3_model.pth: Stores the final model parameters after robust aggregation.
	◦	level3_accuracy_graph.png: Visualizes accuracy during the training process.
	◦	federated_learning_level3.log: Logs accuracy values and detected malicious clients per epoch.

Performance Analysis
The performance of all three levels is analyzed and visualized in a combined graph:
	1	Level 1: Centralized Learning
	◦	Accuracy consistently improves over 20 epochs, reaching 99.36%.
	2	Level 2: Federated Learning
	◦	Accuracy improves steadily over 10 rounds, reaching 98.01%.
	3	Level 3: Robust Federated Learning
	◦	With malicious client detection, accuracy improves over 15 rounds, reaching 98.52%.
	•	Output File:
	◦	accuracy_comparison_graph.png: Shows accuracy comparison across all three levels.
Instructions to Hardcode Data: Update the level1_accuracies, level2_accuracies, and level3_accuracies lists with your performance results in the script to regenerate the graph.

Notes
	•	Ensure that output files are downloaded after execution using the integrated download functionality.
	•	Use GPU acceleration to avoid long training times.
	•	For Level 3, modify thresholds or parameters to experiment with detecting and mitigating malicious clients.
	•	Duplicate or replace output files automatically if they already exist during multiple runs.