Evading obscure communication from spam emails.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Spam email poses significant security risks, including malware and cyber-attacks.
• Current text-centric spam filters often rely on text pre-processing, which may obscure malicious content.
• The increasing volume of spam necessitates advanced filtering techniques.

Purpose of the Study:

• To evaluate the impact of text pre-processing on spam detection accuracy.
• To compare the effectiveness of machine learning (ML) and deep learning (DL) algorithms for email spam classification.
• To investigate the performance of Long Short-Term Memory (LSTM) models in spam filtering.

Main Methods:

• Utilized the Spamassassin corpus for training and testing.
• Implemented and compared ML and DL algorithms, focusing on LSTM.
• Conducted experiments with and without text pre-processing to assess its influence.

Main Results:

• Deep learning models, particularly LSTM, outperformed traditional ML algorithms in spam detection.
• LSTM achieved higher precision, recall, and F1-scores when text pre-processing was omitted.
• The best performance was observed with LSTM without pre-processing: 95.26% precision, 97.18% recall, and 96% F1-score.

Conclusions:

• Text pre-processing can negatively impact the detection of malicious content in spam emails.
• Deep learning approaches, especially LSTM, offer a more robust solution for advanced spam filtering.
• Current ML algorithms struggle with detecting encrypted communication within spam.