Email Classification · Spam / Malicious Detection

End-to-end deep learning pipeline for classifying emails as Ham (legitimate) or Spam / Malicious.

1. Dataset Overview

The Spammalicious Detection dataset contains 100,000 labeled emails collected in 2024, with 30 features spanning content, metadata, security, and behavioral signals. The target variable is binary: 0 = Ham / Legitimate, 1 = Spam / Malicious.

Feature Categories

Content raw_text, subject, body_plain, body_html, language

Metadata from_address, from_domain, to_addresses, date, message_id

Security spf_result, dkim_result, dmarc_result, x_spam_score, received_origin_ip

Behavioral has_attachments, attachment_types, num_urls, num_phone_numbers

2. Model Design

A sequential feedforward neural network was designed for text-based email classification. The model takes 29 encoded features as input and passes them through two dense hidden layers with ReLU activation, a dropout layer for regularization, and a softmax output layer.

Input shape = (29,)

→

Dense units=128 · relu

→

Dropout rate=0.3

→

Dense units=64 · relu

→

Output units=2 · softmax

Note: The design canvas also supports advanced layers such as LSTM, GRU, Bidirectional, BatchNorm, and various pooling layers for more complex architectures.

3. Training Process

Data Splitting

Training Parameters

Optimizer: SGD Learning Rate: 0.010000 Batch Size: 512 Validation Split: 0.20 Epochs: 10 Save Best Model Checkpoint Early Stopping Reduce LR on Plateau

4. Evaluation Results

Test Loss

0.0100

Test Accuracy

0.9987

Final Train Loss

0.0183

Final Val Accuracy

0.9988

Training Progress (per epoch)

Epoch	Loss	Val Loss	Accuracy	Val Accuracy
1	0.4459	0.2354	0.8638	0.9858
2	0.1734	0.0975	0.9849	0.9935
4	0.0505	0.0355	0.9948	0.9971
5	0.0308	0.0402	0.9908	0.9976
6	0.0251	0.0183	0.9908	0.9978
7	0.0183	0.0103	0.9976	0.9981
9	0.0183	0.0118	0.9976	0.9988

Classification Report

5. Summary

✅ Achieved 99.87% test accuracy with consistently low loss and high precision/recall across both classes. The model generalizes well, with validation metrics closely tracking training performance.

The pipeline is ready for deployment in real-world spam filtering, email security research, and multilingual spam pattern detection.

📧 Email Classification Spam / Malicious Detection