Federated Learning Spam Detection System

Overview

A privacy-focused spam detection system built using TensorFlow Lite, FastAPI, Android, and a real-time monitoring dashboard.

The project demonstrates how mobile devices can participate in a collaborative spam detection workflow while minimizing the transfer of sensitive SMS data. Android clients perform local inference using a TensorFlow Lite model and communicate with a centralized FastAPI server that manages model updates, evaluation, and deployment.

---

Features

• Android SMS Spam Detection Client
• TensorFlow Lite On-Device Inference
• FastAPI Backend Server
• Real-Time Monitoring Dashboard
• Model Versioning and Round Tracking
• Accuracy Monitoring
• Azure VM Deployment
• Privacy-Oriented Architecture

---

System Architecture

Android Client
      │
      │ Prediction / Training Data
      ▼
FastAPI Server
      │
      ├── Model Management
      ├── Accuracy Evaluation
      ├── Round Tracking
      └── Model Export
      │
      ▼
TensorFlow Lite Model
      │
      ▼
Redistributed To Clients

      ▲
      │
Streamlit Dashboard
(Real-Time Monitoring)

---

Repository Structure

Spam-Detection/
│
├── android/
│   └── SpamDetector/
│
├── dashboard/
│   └── dashboard.py
│
├── server/
│   └── main.py
│
├── data/
│   └── SMSSpamCollection
│
├── models/
│   ├── round_0.tflite
│   ├── round_0.weights.h5
│   ├── accuracy.json
│   ├── current_round.json
│   └── vocab.json
│
├── build_trainable_model.py
├── train.py
├── requirements.txt
└── README.md

---

Dataset

Dataset Used:

• SMS Spam Collection Dataset

Classes:

• Ham (0)
• Spam (1)

---

Model Architecture

Embedding
      ↓
GlobalAveragePooling1D
      ↓
Dense (ReLU)
      ↓
Dropout
      ↓
Dense (Sigmoid)

---

Training Configuration

Parameter	Value
Vocabulary Size	8000
Sequence Length	100
Embedding Dimension	32
Batch Size	32
Epochs	10

---

Android Client

The Android application performs on-device spam detection using TensorFlow Lite.

Included Assets

round_0.tflite
vocab.json

Technologies

• Kotlin
• TensorFlow Lite
• Android SDK
• Material Components

---

Backend Server

The backend is implemented using FastAPI.

Main Responsibilities

• Serve latest TensorFlow Lite model
• Accept client uploads
• Track training rounds
• Evaluate model accuracy
• Export updated models
• Provide dashboard metrics

Endpoints

GET  /get_model
POST /upload_weights
GET  /metrics
WS   /ws/dashboard

---

Dashboard

The Streamlit dashboard provides:

• Current training round
• Accuracy tracking
• Upload statistics
• Client activity monitoring
• Real-time server events

Start Dashboard

streamlit run dashboard/dashboard.py --server.port 8501

---

Local Setup

Clone Repository

git clone https://github.com/Codeblitz-Ankit/Spam-Detection.git
cd Spam-Detection

Install Dependencies

pip install -r requirements.txt

Train Initial Model

python train.py

Start Backend

cd server
uvicorn main:app --host 0.0.0.0 --port 8000

---

Production Deployment

Development Environment

Lightning AI

Used for:

• Model training
• Experimentation
• Development

Production Environment

Azure Virtual Machine

Used for:

• FastAPI hosting
• Dashboard hosting
• Model distribution
• Monitoring

---

Monitoring

Current Round:

cat models/current_round.json

Accuracy:

cat models/accuracy.json

Generated Models:

ls models/

---

Future Improvements

• Differential Privacy
• Secure Aggregation
• Federated Averaging with True Weight Aggregation
• Multi-Client Simulation
• Transformer-Based Models
• Docker Deployment
• Kubernetes Deployment

---

Author

Ankit Koli

Federated Learning • Android Development • FastAPI • TensorFlow Lite • Machine Learning