[Paper Review] DeepCapture: Image Spam Detection Using Deep Learning and Data Augmentation

Bedeuro Kim, Alsharif Abuadbba+1 more|arXiv (Cornell University)|Jun 16, 2020

Spam and Phishing Detection19 references7 citations

TL;DR

DeepCapture proposes a CNN-XGBoost model with task-specific data augmentation to improve image spam detection, achieving an 88% F1-score—6 percentage points higher than the best existing model (CNN-SVM). The method enhances robustness against unseen spam images by generating diverse training samples through tailored augmentation techniques.

ABSTRACT

Image spam emails are often used to evade text-based spam filters that detect spam emails with their frequently used keywords. In this paper, we propose a new image spam email detection tool called DeepCapture using a convolutional neural network (CNN) model. There have been many efforts to detect image spam emails, but there is a significant performance degrade against entirely new and unseen image spam emails due to overfitting during the training phase. To address this challenging issue, we mainly focus on developing a more robust model to address the overfitting problem. Our key idea is to build a CNN-XGBoost framework consisting of eight layers only with a large number of training samples using data augmentation techniques tailored towards the image spam detection task. To show the feasibility of DeepCapture, we evaluate its performance with publicly available datasets consisting of 6,000 spam and 2,313 non-spam image samples. The experimental results show that DeepCapture is capable of achieving an F1-score of 88%, which has a 6% improvement over the best existing spam detection model CNN-SVM with an F1-score of 82%. Moreover, DeepCapture outperformed existing image spam detection solutions against new and unseen image datasets.

Motivation & Objective

• Address the performance degradation of existing image spam detectors on entirely new and unseen spam datasets.
• Overcome overfitting in deep learning models trained on limited or imbalanced image spam datasets.
• Develop a robust, lightweight CNN-XGBoost framework optimized for image spam detection with minimal hyperparameters.
• Improve generalization by introducing task-specific data augmentation techniques tailored to image spam characteristics.
• Demonstrate superior performance on cross-data training scenarios involving unseen spam datasets.

Proposed method

• Design a deep CNN architecture with eight layers, optimized for image spam detection using hyperparameter tuning (filters, filter size, activation functions, epochs, batch size).
• Integrate XGBoost as a classifier on top of the CNN’s final features to improve decision boundaries and generalization.
• Apply data augmentation techniques specifically tailored to image spam, such as image rotation, scaling, color jittering, and noise injection, to increase training diversity.
• Balance the training dataset by augmenting minority classes (ham) to mitigate class imbalance and improve model robustness.
• Use publicly available datasets (6,000 spam, 2,313 ham) for training and evaluation, with cross-data training scenarios to test generalization.
• Evaluate model performance using F1-score, precision, and recall on both in-domain and out-of-domain (unseen) datasets.

Experimental results

Research questions

• RQ1 Can a CNN-XGBoost model with task-specific data augmentation outperform existing image spam detection models in terms of F1-score?
• RQ2 How does data augmentation impact model generalization on entirely new and unseen image spam datasets?
• RQ3 Does the proposed model maintain high performance in cross-data training scenarios where the test data differs from the training data distribution?
• RQ4 What is the optimal configuration of CNN architecture and hyperparameters for image spam detection with minimal complexity?
• RQ5 To what extent does data augmentation reduce overfitting and improve robustness in image spam detection?

Key findings

• DeepCapture achieves an F1-score of 88%, which represents a 6 percentage point improvement over the best existing model, CNN-SVM (82%).
• In cross-data training scenarios with entirely new and unseen spam datasets, DeepCapture maintains the highest F1-score, outperforming all other classifiers.
• Without data augmentation, all classifiers in cross-data scenarios achieve F1-scores below 40%, highlighting the necessity of augmentation for robustness.
• The integration of data augmentation significantly improves model generalization, especially on unseen spam samples, reducing overfitting.
• The CNN-XGBoost framework with only eight layers achieves superior performance compared to deeper models like the 13-layer CNN-SVM, indicating efficiency and effectiveness.
• The model demonstrates strong performance on real-world image spam datasets, confirming its feasibility and practicality for deployment in spam filtering systems.