Maria Qamar, Nilesh Jain

DOI: https://doi.org/10.5281/zenodo.23397496

The problem of email spam is still extremely topical since the unsolicited bulk messages change in both content and method. This essay examines machine learning-based spam filters on a test email dataset that consists of spam and valid (ham) emails. We use four classifiers, namely Naive Bayes, Support Vector Machine (SVM), Random Forest and Logistic Regression and compare them in terms of measures of accuracy, precision, recall, and F1-score. The models are all high in spam detection accuracy (greater than 90 percent), with SVM being a little better in the overall accuracy (approximately 95.5 percent) and accuracy. Naive Bayes has the best spam recall (around 94.2) but with a more significant number of false positives and a trade-off between spam captures and legitimate email loss. We use extensive text preprocessing (tokenization, stop-word elimination, stemming) and make comparisons of feature representations: word embeddings vs. term frequency-inverse document frequency (TF-IDF). TF-IDF features performed better on this dataset than the average word embeddings on this dataset. In the fight against constantly evolving spam techniques, we consider the use of continuous learning approaches that dynamically update the filter to overcome concept drift, and we consider multimodal approaches that use textual and image data or link analysis to detect material that the purely textual filter fails to detect. As we have found, it is indeed true that using fixed machine learning models is effective in detecting spam and doing it with a high level of accuracy, however next-generation systems do need to constantly learn and make use of several data modalities to serve as a resource to robust and adaptive spam filtering in dynamic real-life scenarios.

spam detection; continuous learning; concept drift; multimodal; text classification; email security

We have in this work proposed a complete work on spam email detection based on machine learning, and have mentioned extensions of continuous learning and multimodal filtering that are essential in next-generation spam defenses. On a benchmark set of 4,601 emails, we trained four classifiers (Naive Bayes, SVM, Random Forest, Logistic Regression) on extensive text preprocessing and 2 feature extraction methods (TF-IDF and word embeddings). The content-based filters were effective in the controlled environment as all the models were accurate on the test data (93-95%). The SVM model had the highest overall performance with a 95.5 percent accuracy with a 95.3 percent precision and 93.0 percent recall where Naive Bayes was the most aggressive in spam-catching with the highest recall (94.2 percent) at the cost of the lowest precision (89 percent). Through analysing the errors and feature weights we discovered that both models have a characteristic bias: some will focus on false negative reduction (such as NB) and others false positive reduction (such as SVM). A combination or mixed method may possibly give the best of both, but with our findings even a basic Logistic Regression or a Random Forest was already giving a desirable level of precision and recall of around 92-93%. Notably, our experiments confirmed that the representation of features is important. TF-IDF sparse features performed better than averaged word embeddings at this task, presumably because certain keywords (e.g., "winner", "free" and "offer") contain important signal of spam, which gets washed out in an averaged embedding. This implies that at least in the classical models it is advantageous to retain those granular keyword features. More advanced models of the neural network might have an advantage of using embeddings without dropped signal on key-words, but it would require greater amounts of training data and design. In addition to the snap-on examination, another theme of this paper is that spam detection requires continuous learning. Spam does not stand up: spammers keep adding to tricks (topics, obfuscations, languages) to circumnavigate the filters. A model that has been trained will always grow out of fashion. We talked about the realization of an adaptive system: by means of incremental learning online (training the model with every new email label or every periodically training with new data), by means of drift detection techniques to be aware when the system is performing poorly, and by means of user feedback mechanisms to include human judgments. The spam filter keeps up with the changing but not changing spam environment by learning continuously, which is how it is possible to close the concept drift problem that has been discussed in recent studies. Continuous learning also entails protection against forgetting of older spam experience – our discussion referred to the ensemble tactics and data remembrance approaches in order to make sure that, e.g., the revival of an old spam strategy can still be identified by the filter. These measures would enable a spam detector engine to keep up with the spammers, and substantially increase its useful life without necessarily having to be retrained entirely

1.Karim A, Azam S, Shanmugam B, Kannoorpatti K, Alazab M. A comprehensive survey for intelligent spam email detection. IEEE Access. 2019;7:168261-168295.
2.Cranor LF, LaMacchia BA. Spam! Commun ACM. 1998;41(8):74-83.
3.Goodman J, Heckerman D, Rounthwaite R. Stopping spam. Sci Am. 2005;292(4):42-49.
4.Abu-Nimeh S, Nappa D, Wang X, Nair S. A comparison of machine learning techniques for phishing detection. Proc eCrime Researchers Summit. 2007;pp. 60-69.
5.Blanzieri E, Bryl A. A survey of learning-based techniques of email spam filtering. Artif Intell Rev. 2008;29(1):63-92.
6.Sahami M, Dumais S, Heckerman D, Horvitz E. A Bayesian approach to filtering junk e-mail. Proc AAAI Workshop on Learning for Text Categorization. 1998;pp. 55-62.
7.Guzella TS, Caminhas WM. A review of machine learning approaches to spam filtering. Expert Syst Appl. 2009;36(7):10206-10222.
8.Kolter JZ, Maloof MA. Dynamic weighted majority: an ensemble method for drifting concepts. J Mach Learn Res. 2007;8:2755-2790.
9.Nosrati M, Pour MS. Email spam filtering using dynamic weighted majority with concept drift detection. Int J Comput Sci Inf Secur. 2011;9(9):1-7.
10.Drucker H, Wu D, Vapnik V. Support vector machines for spam categorization. IEEE Trans Neural Netw. 1999;10(5):1048-1054.
11.Carreras X, Márquez L. Boosting trees for anti-spam email filtering. Proc RANLP. 2001;pp. 58-64.
12.AbdulNabi ZM, Yaseen QS. A BERT-based approach for email spam detection. Int J Adv Comput Sci Appl. 2021;12(10):400-407.
13.Dedeturk M, Akay B. A content-based deep learning approach for email spam filtering. IEEE Access. 2020;8:54033-54045.
14.Lowd D, Meek C. Good word attacks on statistical spam filters. Proc CEAS. 2005;pp. 125-132.
15.Biggio B, Corona I, Maiorca D, Nelson B, Šrndić N, Laskov P, et al. Evasion attacks against machine learning at test time. Proc ECML PKDD. 2014;pp. 387-402.
16.Dua D, Graff C. UCI Machine Learning Repository: Spambase [Data set]. Irvine, CA: University of California, School of ICS; 2019.
17.Bird S, Klein E, Loper E. Natural Language Processing with Python. O’Reilly Media; 2009.
18.Porter MF. An algorithm for suffix stripping. Program. 1980;14(3):130-137.
19.Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20(3):273-297.
20.Jáñez-Martino F, Alaiz-Rodríguez R, González-Castro V, Fidalgo E, Alegre E. A review of spam email detection: analysis of spammer strategies and the dataset shift problem. Artif Intell Rev. 2023;56(2):1145-1173.
21.Goodman J, Yin H. Online discriminative spam filter training. Proc CEAS. 2006.
22.Shah F, Xia C. Investigating the effectiveness of Word2Vec for spam detection with machine learning algorithms. Int J Intell Syst Appl Eng. 2022;10(1):15-22.

Maria Qamar, Nilesh Jain (2023). Continuous Learning and Multimodal Techniques for Next-Generation Spam Detection Systems. International Journal of Computer Techniques, 10(6), 7–23. ISSN: 2394-2231. DOI: https://doi.org/10.5281/zenodo.23397496