What question did this study set out to answer?

The research aims to develop a spam detection model that effectively identifies spam while reducing false positives and adapting to new spam techniques.

April 4, 2026Journal of university of Anbar for pure science0 citationsOpen Access

Enhanced email spam filtering using spectral clustering and deep neural networks

SASahar AhmedUniversity of Anbar HSHadeel M. SalehUniversity of Anbar SASameer Alani

Key Points

The research aims to develop a spam detection model that effectively identifies spam while reducing false positives and adapting to new spam techniques.
Developed SCDNN combining spectral clustering and deep neural networks for spam detection.
Partitioned email corpus based on centroids from spectral clustering methods.
Utilized deep neural networks to assess spam likelihood by comparing training and test data points.
SCDNN achieved 96.3% accuracy on the Enron Email Corpus, outperforming CNNs at 94.7%.
On the Spam Assassin dataset, SCDNN scored 91.6% accuracy compared to LSTM networks at 93.0%.
SCDNN showed better generalization against evolving spam tactics across various datasets.

Abstract

The vast growth in email communications has resulted in an overwhelming increase in unsolicited email, making it challenging to filter properly. This study proposes a novel spam detection approach called SCDNN, which adapts spectral clustering and deep neural networks for precise spam identification. The aim of the study is to engineer an enhanced filtering model capable of delivering precise spam detection while minimizing false alarms and strengthening the system’s adaptability against evolving deceptive techniques. Firstly, SCDNN partitions the corpus according to neighborhoods centered around clustering centroids. Next, deep neural networks estimate spam by gauging distances between test and training data points using likelihood metrics. Experiments on several email datasets, such as the expansive Enron collection, Spam Assassin corpus, and UCI Mail Spam benchmark, indicate that SCDNN outperforms traditional techniques like backpropagation neural networks, support vector machines, random forests, and Bayesian logic with respect to precision, recall, mean accuracy, and F1 score. For instance, SCDNN achieved 96.3% accuracy on the Enron Email Corpus versus 94.7% for convolutional neural networks, and 91.6% accuracy on Spam Assassin compared to 93.0% for long short-term memory networks. The findings suggest that SCDNN exhibits superior generalization against ever-shifting spam tactics across diverse evaluation sets. It is this scalable methodology that enables large-scale filtering analysis and holds great promise for widespread application. SCDNN can mitigate the spam load, enhance user efficiency, and lower the risks associated with spam, including phishing and malware.

Demander à l'IA

Bookmark

View Full Paper