SMS Scam Detection Application Based on Optical Character Recognition for Image Data Using Unsupervised and Deep...

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Unsolicited text messages, known as smishing, and data irregularities pose significant security challenges.
• Existing spam detection methods often overlook the complex interplay of text, images, and context in smishing attacks.
• There is a need for advanced models that can effectively analyze these multifaceted relationships.

Purpose of the Study:

• To develop and evaluate a sophisticated model for detecting smishing messages.
• To explore the effectiveness of combining traditional machine learning and deep learning techniques for spam detection.
• To analyze the relationships between words, images, and contextual factors in identifying smishing.

Main Methods:

• Merged a UCI spam dataset with real-world spam data, utilizing Optical Character Recognition (OCR) for image analysis.
• Employed traditional machine learning models (K-means, NMF, GMM) with feature extraction (TF-IDF, PCA).
• Utilized deep learning models (RNN-Flatten, LSTM, Bi-LSTM) to capture sequential dependencies and contextual nuances.

Main Results:

• The K-means feature extraction with vectorizer achieved 91.01% accuracy.
• The KNN-Flatten model demonstrated superior performance, reaching 94.13% accuracy.
• Comparative analysis highlighted the strengths of both machine learning and deep learning approaches in spam detection.

Conclusions:

• The KNN-Flatten model shows significant potential for real-time smishing detection due to its high accuracy.
• While effective, the computational complexity of advanced models like KNN-Flatten may impact large-scale deployment.
• K-means with vectorizer offers good accuracy but may require continuous retraining to adapt to evolving smishing tactics.