Indic PrivacyFL

A robust Federated Learning framework designed for categorizing Multimodal Indic Educational Data across 7 languages and 8 subject domains without compromising student data privacy.

Project Objectives

1. Data & Challenge Characterization: Analyze multimodal Indic educational data (text, speech, image, video) and cross-lingual code-mixed language patterns.
2. Robust Multimodal Architecture: Implement a multi-encoder network using state-of-the-art foundation models with a Self-Attention Fusion mechanism that degrades gracefully when client modalities are missing.
3. Privacy-Preserving Prototype: Deploy a distributed Federated Learning simulation using Differential Privacy (Opacus DP) to guarantee user privacy at the node level.
4. Evaluation & Benchmarking: Measure performance across 8 educational subjects, comparing Centralized baselines against Federated implementations.

Quick Start (Demo Mode)

For a rapid, zero-download test run of the architecture locally:

# 1. Create and activate a virtual environment
python -m venv venv
# Windows: venv\Scripts\activate
# Linux/macOS: source venv/bin/activate

# 2. Install dependencies
pip install -r requirements.txt

# 3. Generate instant demo data (50 samples)
python -m src.data.download_data --demo

# 4. Run the fast federated simulation demo
python -m src.fl.run_simulation --demo

️ Environment Setup

1. Requirements

• Python 3.10+
• PyTorch 2.1+ (CUDA recommended for full training)

2. Installation

python -m venv venv

# Windows
venv\Scripts\activate

# macOS/Linux
source venv/bin/activate

pip install -r requirements.txt

Data Acquisition

The framework supports multiple data acquisition modes depending on your hardware limits and network access.

A. Demo Version (Instant)

Generates 50 minimal synthetic records for instant pipeline testing. No downloads required.

python -m src.data.download_data --demo

B. Sample Version (Rapid Local Prototyping)

Generates lightweight multimodal sample records (500 samples/language).

python -m src.data.download_data --sample

C. Full Version (HuggingFace Stream)

Attempts to stream the full multimodal educational corpora (AI4Bharat Sangraha, IndicVoices-R) via HuggingFace. If authentication or quotas fail, it falls back to extended synthetic generation with a printed warning.

python -m src.data.download_data --full

Running the Pipeline

1. Centralized Training Baseline

Train a centralized baseline model to evaluate the upper-bound accuracy without federation constraints.

python -m src.fl.run_centralized

2. Federated Learning Simulation

Run the distributed federated learning network across multiple clients using Flower (flwr).

# Standard Federated Simulation (No DP)
python -m src.fl.run_simulation

# Federated Simulation with Differential Privacy (DP)
python -m src.fl.run_simulation --dp-enabled

3. Interpreting Results

To generate comparison tables and plots across Centralized vs. Federated results:

python -m src.eval.compare_results

Note: Results are saved to the results/ directory as comparison_plot.png and comparison_table.csv.

️ Command-Line Flags Overview

run_simulation.py Flags

Project Structure

IndicPrivacy-FL/
├── configs/            # YAML configuration files (default.yaml)
├── data/               # Raw and processed multimodal data
├── DIAGRAMS/           # Architectural diagrams
├── docs/               # Technical architecture documentation
├── reports/            # Output reports and evaluation summaries
├── results/            # Generated metrics, tables, and plots
├── src/
│   ├── data/           # Dataset loaders and partitioners
│   ├── eval/           # Evaluation and comparison scripts
│   ├── fl/             # Federated clients, server logic, strategies
│   └── models/         # Modality encoders, fusion mechanisms
└── tests/              # Unit tests for core components

️ Architecture Documentation

For technical implementation details, please refer to the documentation:

• Data Pipeline Architecture
• Multimodal Model Architecture
• Federated Learning System Architecture

Running Tests

To verify the integrity of the core components (encoders, fusion, and model):

python -m pytest tests/ -v

️ Differential Privacy Configuration

Configured via configs/default.yaml.

Datasets & Technologies

Please see TOOLLIST.md for the complete list of target datasets, open-source libraries, and foundational models (IndicBERT, Wav2Vec, ViT) powering this framework.