Email Spam Detection Using Machine Learning
C. Venkata Swamy1, Nagineni Sreehari2, Shaik Moulali2, Mulla Mohammed Saleem2,
Shaik Showkath Ali2 and S. Nasir Hussain2
1Department of Computer Science and (AI‑ML), Santhiram Engineering College, Nandyal‑518501, Andhra Pradesh, India
2
Department of Computer Science and Design, Santhiram Engineering College, Nandyal‑518501, Andhra Pradesh, India
{[Link], harisrihari758, moulalishaikmj, saleemmohamed0786, shaikshowkathali7777, nasirhussainsunkari8688}
@[Link]
Keywords: Spam‑Detection, ML, Natural Language Processor, DL, Classification Models, Email Security, Feature
Extraction.
Abstract: Although email is an integral element of daily communication, increasing amounts of spam email not only
serves as a potential threat to security but also harms efficiency. Traditional spam filters can’t keep pace with
new spamming techniques, making them ineffective over time. Instead of relying on fixed heuristics, machine
learning provides a more adaptive solution to spam detection. We use linear classifiers, such as Naïve Bayes
and Support Vector Machines (SVM), demographic models like Decision Trees and deep-learning models to
achieve better spam detection accuracy in this work. We can apply various natural language processing (NLP)
methods like also Tokenization, stop word removal, TF-IDF and word embedding, to improve the model’s
understanding capacity of the email content. We demonstrate through experimentation, that the performance
of machine learning based spam filters reduces false positives and achieves higher classification accuracy.
Since these models learn from data again, they can quickly adapt to the evolving spam techniques and
therefore, are a reliable solution in modern email security. Further developments can supplement real-time
learning, hybrid models, and deep learning to significantly improve email spam detection systems.
1 INTRODUCTION 1.1 Motivation
Email spam, or junk mail, is the use of messaging The study of ml techniques in email spam-detection
systems to send unsolicited messages (spam), usually is the growing demand of correct and fast filter. The
advertising for some product, service, or other hybrid framework of advanced classifiers with rule-
activity. These types of mails have become increasing based filtering not only improves email filtering
numbers on the internet in the past 10 years and now performance and accuracy but also enhances the
have become a notable nuisance on the internet Spam user's email experience. This combination thus
emails take up storage space and waste time, they provides for a highly flexible, accurate and intelligent
delay the delivery of messages. Despite automatic detection approach, successfully limiting false
email filtering being one of the most effective spam alarms while improving spam finding. This technique
detection methods available, spammers have come up helps to make the email communication landscape
with many smart ways to avoid such filtering systems. more secure and trustworthy by addressing the
In the past, spam emails were mostly filtered out by changing landscape of spam methods.
blacklisting specific email addresses. But as
spammers generate new email domains, this method 1.2 Objectives
has become increasingly ineffective. Spam detection
has gained new interest with the development of An approach to build a multi spam-detection system
various ml techniques. There are several popular which utilizes both traditional rule-based filtering and
methods for spam filtering are text analysis, using ml classifiers (SVM, ANN, and XG Boost).
blacklists and whitelists based on domain names, and Step1: to use human language processing techniques
network-based methods. by making the email preprocessing [data cleaning and
feature extraction].
807
Swamy, C. V., Sreehari, N., Moulali, S., Saleem, M. M., Ali, S. S. and Hussain, S. N.
Email Spam Detection Using Machine Learning.
DOI: 10.5220/0013873400004919
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages
807-811
ISBN: 978-989-758-777-1
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,
COMMUNICATION, AND COMPUTING TECHNOLOGIES
Learn from data until October 2023, To Investigate Boost have been employed to improve classification
model's ability to adapt to changing threats and performance through the aggregation of several
discover factors for improvement such as model outputs (H. Drucker et al., 1999).
transformer-based deep learning model and real-time
learning. 2.4 Gaps in Existing Research
Despite getting much better, existing spam detection
2 LITERATURE REVIEW systems still face some obstacles:
• Challenges to identify advanced phishing or
In this section discussed the progress of the being part of adversarial attacks that send
technology of email spam-detection, representing spam
both the conventional and current ways. • False-positive rates too high, causing
legitimate emails to be classified as spam.
2.1 Traditional Spam Filtering • Scalability issues in processing large and
Methods changing email datasets.
One of the most common methods was Bayesian A hybrid spam detection adaptability and accuracy
filtering, a probabilistic method that determines the this review on the literature
probability of an email being spam based on word
frequency (Androutsopoulos et al., 2000). Although
these approaches worked well in punishing spam 3 METHODOLOGY
above based purely on the presence of specific
keywords, they struggled to keep up with changing In the methodology section, we describe the system
tactics spammer used (for example, obfuscation design, the data processing pipeline, the integration
techniques, or adversarial attacks). of machine learning and evaluation approach.
2.2 ML in Spam-Detections 3.1 System Architecture
Machine learning spam-detection ushered in a new The proposed system is composed of four major
era by allowing models to learn from patterns in the modules:
data, rather than hard-coding rules. Classic machine
learning models like SVM, Decision Trees and 3.1.1 Email Data Collection
Naïve Bayes have been used to classify spam and ham
based on the features extracted from the data such as A data collection module gathers email data from
word frequency, metadata and header 2. Despite publicly available datasets (such as Enron Spam
having shown improved performance over the Datasets) and real-world email traffic. The dataset
previous rule-based approaches, these models suffer comprises two spams and legitimate (hams) emails,
from evolving spam types and the scale of these email ensuring a balanced and diverse corpus.
datasets.
3.1.2 Preprocessing and Feature
2.3 Deep Learning and Advanced Engineering
Techniques
After data collection, preprocessing is performed to
extract meaningful features:
Spam Detection improves with the recent
advancement in deep learning Neural-network has
been used for analysing complex patterns in the text Tokenization and Stop-word Removal:
and metadata of emails, especially Convolutional The email text is broken into tokens,
Neural Networks (CNNs) and Recurrent Neural and unique words (e.g., "and")
Networks (RNNs) (W. S. Yerazunis, 2004). More removed enhance relevant content
recent transformer-based models like BERT, were extraction.
designed to enhance context awareness and better Stemmings and Lemmatizations:
identify spam emails (J. Goodman, 2005). In addition, Words are reduced to their root forms
ensemble methods such as Random Forests and XG for text normalization.
808
Email Spam Detection Using Machine Learning
Metadata Analysis: Additional An input layer representing extracted
features, such as sender reputation, features.
frequency of links, and email structure, One or more hidden layers
are extracted. capturing relationships within the
data.
3.1.3 Classification Engine An output layer classifying emails
as spam or ham.
This module interprets email contents and metadata • XG Boost: XG Boost efficiently captures
to classify them as spam or legitimate. Various feature interactions and mitigates
machine learning models are applied to improve overfitting. The final classification score is
classification accuracy. computed as:
3.1.4 Spam Filtering Module: Ci = α*SVMi + β*ANNi + γ*XGBoosting (1)
Detected spam emails are flagged and either moved
to the spam folder or discarded. The system where Ci is the final classification score, and α, β,
continuously learns from new emails to enhance and γ are weight parameters optimized during
detection performance. Figure 1 shows System training.
Architecture for Machine Learning Based -Spam where Ri is the final ranking score, PRi is the
Detection. PageRank score, MLi is the machine learning output,
and α and β are weight parameters optimized during
training. Figure 2 shows DF Diagram.
Figure 1: System architecture for machine learning based -
spam detection.
Figure 2: DF diagram.
3.2 Machine Learning Integration
3.3 Data Collection and Preprocessing
The hybrid spam detection model employs three
primary ml technique: The dataset is curated through publicly available
sources and real-time email monitoring:
• Support Vector Machines (SVM): It is • Dataset Composition: Over 100,000
SVM is utilized to classify emails based on emails are collected, consisting of both
textual features. Use non-linears kernel (e.g. spams and non-spams emails across various
polynomial), it handles complex decision domains.
boundaries, improving spam classification • Pre-processing Pipeline:
performance.
• Data Cleaning: Removal of HTML
• Artificials Neural Networks (ANN): A tags, special characters, and non-
multi-layers perceptrons (MLP) is text elements.
implemented to the detect intricate patterns
• Text Normalization: Tokenization,
in emails:
stemmings, and lemmatizations.
809
ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,
COMMUNICATION, AND COMPUTING TECHNOLOGIES
• FEATURES Vectorization: 4.2 Comparative Analysis
Conversion of text to numericals
vectors using TF-IDF and word
embeddings.
3.4 System Evaluation and Testing
The system is evaluated using both quantitative and
qualitative measures:
• Quantitative Metrics:
• Confusion Matrix Analysis:
Evaluates false positive and false
negative rates.
• Qualitative User Feedback:
• User Surveys: Gather user
feedback on spam detection
accuracy and usability.
• Spam Reduction Impact: Measures Figure 3: Comparative performance based on key
the reduction of spam emails in evaluation metrics.
user inboxes over time.
• Regression Analysis: Multiple regression Figure 3 of the Comparative analysis the graph
models (using tools like IBM SPSS) show is Comparative Performance Based on Key
determine the influence of various features Evaluation Metrics.
on classification accuracy.
5 DISCUSSION
4 RESULT AND ANALYSIS
This section details the experimental findings from 5.1 Adaptive Learning and Relevance
implementing the hybrid search engine.
This enables the system to adjust to new patterns of
4.1 Qualitative Evaluation spam and changing underlying structures of emails.
Such models learn from new data and iterate on the
• User Feedback: classification, thus increasing accuracy as more data
Through a study with 200 email customers, is captured over time. Training reported higher
this machine learning-based spam filter precisions, recalls, and F1-scores metrics, thus
dramatically improves the accuracy of email enhanced system performance, decreasing false
filtering. Users saw less spam in their positive and false negative and allowing more
inboxes and were able to identify unwanted accurate and trustful spam detection.
messages.
• Usability Testing: 5.2 User Experience and Security:
Consensus from focus group discussions
indicates that the spam filtering system is The system leads to increased email security as it
intuitive, non-intrusive, and blends detects spam and minimizes exposure to phishing
seamlessly into the email platforms. The emails. Increased trust isn't just good for you, it helps
performance of the automated classification to build a better user experience by ensuring that users
was found to be satisfactory by the users, recognize genuine promotional emails from harmful
and thus reducing the need for manual spam.
intervention. They also observed how the
system automatically adopts to new spam 5.3 Limitations
trends, retaining high accuracy over time.
High computational cost of training deep learning
models and tendencies of training data for bias that
810
Email Spam Detection Using Machine Learning
might have an effect on generalization are some of the
challenges. Moreover, the real time filtering would
add latencies and need optimizations for a production
scale deployment such as in enterprise email
services.
6 CONCLUSIONS
Machine learning has proven to be an effective
approach for detecting and filtering spam emails,
significantly improving classification accuracy
compared to traditional rule-based methods. By
utilizing advanced algorithms such as Naïve Bayes,
Support Vector Machines (SVM), Decision Trees,
Random Forest, and deep learning models, spam
detection systems can efficiently distinguish between
spam and legitimate emails.
REFERENCES
H. Drucker, D. Wu, and V. N. Vapnik, “Support vector
machines for spam categorization,” IEEE Trans. Neural
Networks, vol. 10, no. 5, pp. 1048–1054, Sep. 1999.
I. Androutsopoulos, G. Paliouras, V. Karkaletsis, G. Sakkis,
C. D. Spyropoulos, and P. Stamatopoulos, “Learning to
filter spam e-mail: A comparison of a Naive Bayesian
and a memory-based approach,” Intelligent Systems,
vol. 37, no. 4, pp. 415–429, 2000.
J. Goodman, “Spam filtering: Bayesian and beyond,” in
Proc. 2nd Conf. Email Anti-Spam (CEAS), 2005, pp.
1–9.
S. Islam, T. K. Ghosh, and M. S. Rahman, “A deep
learning-based approach for email spam classification
using hybrid CNN-LSTM model,” in Proc. IEEE Int.
Conf. Signal Process. Inf. Comput. Appl. (SPICA),
2021, pp. 203–208.
T. A. Almeida, J. M. Gómez Hidalgo, and A. Yamakami,
“Spam filtering: How the dimensionality reduction
affects the accuracy of Naïve Bayes classifiers,”
Journal of Information Processing & Management, vol.
47, no. 5, pp. 654–664, 2011.
W. S. Yerazunis, “Sparse binary polynomial hashing for
spam filtering,” in Proc. ACM Conf. Email Anti-Spam
(CEAS), Mountain View, CA, USA, 2004, pp. 1–8.
811
