Spam Filtering in the Modern Era: A Review of Machine Learning,

Deep Learning, and System Comparisons

Xinyi Wang

School of Electrical Engineering and Artificial Intelligence, Xiamen University Malaysia,

Jalan Sunsuria, Bandar Sunsuria, Malaysia

Keywords: Spam Filtering, Machine Learning, Deep Learning, Navie Bayes, Support Vector Machines, Random Forest,

BERT, Convolutional Neural Networks, Concept Drift, LLM-Modified Spam.

Abstract: Spam poses a great challenge to Internet users, threatening their productivity, data security, and overall user

experience. This review examines the current state of spam filtering systems, particularly those using Machine

Learning (ML) and Deep Learning (DL) models. A comparative analysis of models such as Naive Bayes,

Support Vector Machines (SVM), Random Forest (RF), Bidirectional Encoder Representations from

Transformers (BERT), and Convolutional Neural Networks (CNNs) is conducted to assess their effectiveness

and limitations in distinguishing spam from legitimate (ham) emails. Additionally, this study highlights

emerging challenges, including Concept Drift and Large Language Model (LLM)-modified Spam, which

necessitate the development of more adaptive and intelligent filtering solutions. The future trends indicate

that the spam filtering system will gradually shift to Artificial Intelligence (AI)-driven automation and

personalized, with lifelong learning models that continuously learn from new data leading the way. Such

advancements are essential to counter sophisticated, hard-to-detect spam, like sophisticated LLM-crafted

spam, like those crafted by large language models, ensuring a safer and more personalized email experience

for users.

1 INTRODUCTION

In the digital age, spam has become a common

problem, filling the inbox and posing a major threat

to the safety and productivity of users. Spam usually

contains phishing links, malware or misleading

advertisements aimed at deceiving recipients or

promoting unsolicited products. And as spammers

continue to enhance their tactics, it is becoming more

difficult for traditional filtering methods to keep up

with the changing pattern of spam.

This article explores the development of spam

filtering technology, focusing on the transition from

early rule-based systems, such as SpamAssassin, to

contemporary machine learning and deep learning

methods. By comparing models, such as Navie Bayes,

Support Vector Machine (SVM), Random Forest and

Bidirectional Encoder Representations from

Transformers (BERT), a comprehensive overview of

the advantages and disadvantages of current spam

filtering technologies is provided. In addition, this

study addresses emerging challenges and future

trends in spam detection, underscoring the necessity

of adaptability and innovation in combating spam.

2 DEFINITION AND

CHARACTERISTICS OF SPAM

EMAILS

Spam emails are the unsolicited messages which are

typically sent in bulk to a wide range of users. These

emails generally contain phishing links, malware or

unwanted advertisements, aiming to mislead

recipients or promote products (Aragão Ferreira,

Oliveira, Kuehne, Moreira, & Carpinteiro, 2021;

Ghosh & Senthilrajan, 2023). Spam emails not only

result in the loss of productivity but also leads to the

exposure of inappropriate content to inappropriate

audiences (Aragão et al., 2021).

There are three main types of spam emails. The

first category comprises commercial advertisement

emails, which are typically filled with promotional

messages designed to sell products or services. The

Wang and X.

Spam Filtering in the Modern Era: A Review of Machine Learning, Deep Learning, and System Comparisons.

DOI: 10.5220/0013526000004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 451-458

ISBN: 978-989-758-754-2

451

second, phishing emails, impersonates reliable

sources in an attempt to obtain credit card numbers or

passwords. Finally, malware emails usually carry

viruses or other harmful programs that can damage

your device and compromise your data security once

opened. (Kaddoura, Alfandi, & Dahmani, 2020).

Moreover, these emails often exhibit some

common notable characteristics. First of all, they

frequently use some specific keywords and phrases,

which are eye-catching and likely related to

promotions, investment opportunities or impractical

promises, such as “free”, “urgent”, and “limited time

offer”. Secondly, many spam emails contain images,

and such images may hide harmful content or induce

the user to click on the link. Besides, some spam

emails will also contain suspicious links leading to

fake websites or downloads. And additionally,

attachments is also a common feature of spam, which

may contain viruses or other malware.

3 DEVELOPMENT OF SPAM

FILTERING TECHNIQUES

Spam filtering technologies have experienced a

remarkable development in the past few decades.

However, with the popularity of internet and the

widespread application of emails, the quantity and

types of spam emails are continuously increasing. As

a consequence, developing effective spam filtering

technology is becoming more and more important.

3.1 Early Filtering Techniques (Rule-

Based Filtering)

The earliest spam filtering method was mainly a rule-

based system, such as SpamAssassin. These systems

applied more than 700 tests and assign the scores of

spam emails based on a series of predefined rules. For

example, the system would check for specific

keywords in emails, the sender's reputation, and

header information. Specifically, they combined

technologies, including Bayesian filtering,

blacklisting, and Domain Keys Identified Mail

(DKIM) verification. This method can identify spam

by analyzing the content and source of the email, so

as to effectively filter out most of the obvious spam

(Elliott, 2023). However, with the continuous

evolution of spam methods, the traditional rule-based

filtering technology has gradually become less

flexible and accurate.

Although these early filtering systems were

effective at the time, traditional rule-based filtering

techniques have become increasingly limited in

flexibility and accuracy. This is because spammers

began using more complex strategies, such as

disguising email content and altering sender

addresses, which made it difficult for many filtering

systems to effectively identify spam. Consequently,

relying solely on rule-based methods has proven

inadequate for addressing the increasingly

sophisticated challenges posed by spam (Mohammad,

2020).

3.2 Transition to Modern Filtering

Methods

As spammers are increasingly better at evading

simple filtering rules, researchers began to turn to

study more complex modern filtering methods. The

emergence of machine learning has provided a new

solution for spam filtering. Machine learning

algorithms, such as Naive Bayes and Support Vector

Machines (SVM), can train models on large data sets

and identify complicated characteristics of spam

emails. Besides, these algorithms can learn how to

distinguish between spam and ham emails by

analyzing historical mail data, in order to improve the

accuracy of filtering. For instance, the Naive Bayes

algorithm uses Bayes' theorem to analyze the

frequency and probability of keywords in historical

email data, learning how to effectively distinguish

between spam and ham emails (Feng, Sun, Zhang,

Cao, & Yang, 2016). And on the other hand, Support

Vector Machines work by finding the optimal

hyperplane in feature space to maximally separate

spam from ham emails (Ghosh & Senthilrajan, 2023).

These algorithms can continuously learn and

optimize through historical data analysis, thereby

improving filtering accuracy.

3.3 Current Research Trends (Machine

Learning and Deep Learning

Approaches)

In recent years, deep learning approaches have gained

wide attention on detecting spam emails. Such deep

learning models, like BERT (Bidirectional Encoder

Representations from Transformers), can greatly

improve the accuracy of spam classification through

using attention mechanism and context understanding.

And these models can effectively capture the subtle

nuances in emails, enabling them to deal with diverse

spam characteristics (Devlin, Chang, Lee, &

Toutanova, 2019). The current research trend focuses

on optimizing these algorithms to improve the

efficiency and accuracy of filtering, and respond to

constantly changing spam strategies. For example,

researchers are exploring how to combine the

DAML 2024 - International Conference on Data Analysis and Machine Learning

452

advantages of multiple models to enhance overall

detection capability. Additionally, leveraging natural

language processing techniques to further understand

the context and semantics of email content has also

emerged as a new research direction (Vaswani,

Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, &

Polosukhin, 2017).

In conclusion, the development process of spam

filtering technology illustrates the transformation

from simple rule-based systems to complex

intelligent algorithms. With ongoing technological

advancements, spam filtering systems are progressing

toward higher efficiency and intelligence, ultimately

aiming to provide users with a safer and cleaner email

environment.

4 MAIN METHODS OF SPAM

FILTERING

4.1 Content-Based Filtering

In the content-based filtering methods, the core is the

comprehensive analysis of email content, so as to

capture the potential characteristics of spam emails.

This method usually evaluate the text, links and

attachments in emails to determine whether they meet

the specific spam characteristics. Through the

integrated application of technologies, such as

Natural Language Processing (NLP), keyword

detection, pattern matching and context analysis, this

method can effectively identify and filter spam emails

(Saidani, Adi, & Allili, 2020). Outlined below are

several key technologies and their implementations

within this approach.

4.1.1 Keyword Detection

NLP technology is the basis of keyword detection. It

can analyze words and phrases in email text,

identifying the content that is highly relevant to spam

emails. For example, promotional and eyecatching

words, such as “free”, “discount” and “winning” are

often used in spam emails. To support detection, the

filtering system establishes a dataset containing

commonly used spam words (Junnarkar, Adhikari,

Fagania, Chimurkar, & Karia, 2021). Consequently,

the system would analyze the keywords in the mail to

determine whether it is likely to be a spam. Besides,

this process can be enhanced with machine learning

algorithms, which means that the system could learn

and update emerging spam words and expressions to

maintain the efficiency of filter (Roumeliotis,

Tselikas, & Nasiopoulos, 2024).

4.1.2 Pattern Matching Technology

Pattern matching technology is mainly used to

identify the unusual formats or structures in emails.

Spammers typically adopt tactics which can evade

conventional detection, such as using irregular type

setting, abnormal coding or special characters. But by

analyzing the metadata and header information of

emails, pattern matching technology can detect

improperly formatted email headers, non-standard

character encodings and suspicious embedded links

(Beaman & Isah, 2021). In addition, pattern matching

can also recognize specific HyperText Markup

Language (HTML) and Cascading Style Sheets (CSS)

codes, which are often used by spammers to cover up

malicious links or hide actual content (Zych, 2020).

4.1.3 Contextual Analysis and Semantic

Understanding

In content-based filtering, simple keyword detection

is insufficient to cope with all spam. Contextual

analysis will further reduce misjudgment by

evaluating the overall semantics of the email and

considering the meaning of words in different

contexts. For example, the word "free" may be a sign

of spam in advertising emails, but in customer support

emails it is reasonable. Supported by NLP technology,

the filtering system can deeply understand the

grammar and semantic structure of the message and

decide whether it is spam according to the context (Si,

Wu, Tang, Zhang, & Wosik, 2024). Thus, this

semantic analysis technology effectively improves

the accuracy and intelligence of the system,

minimizing the risk of blocking legitimate emails.

4.1.4 Multi-layer Test Mechanism of

SpamAssassin

SpamAssassin is a classic representative of content-

based spam filters. It uses a multi-level testing

mechanism to identify spam emails by scanning

various suspicious features in emails. SpamAssassin

will firstly detect common spam keywords, and also

analyze the format and structure of the email to

identify non-standard mail headers or abnormal

character encodings (Chu, Hsu, Sheu, Yang, & Lee,

2020). In addition, SpamAssassin will detect the links

in the email to identify potential malicious or

suspicious links. And if certain feature in the email

match known spam patterns, then the system will

score the email. When the cumulative score exceeds

the preset threshold, the email will be labelled as

spam. This scoring mechanism enables

SpamAssassin to handle different types of emails

Spam Filtering in the Modern Era: A Review of Machine Learning, Deep Learning, and System Comparisons

453

more flexibly, reducing the possibility of

misjudgment (Elliott, 2023).

4.2 Behavior-Based Filtering

Behavior-based spam filtering not only relies on the

analysis of email content, but also provides more

accurate filtering by monitoring and evaluating

email-related behavior. This method identifies

potential spam emails by observing how users interact

with emails and the sender's behavior history.

Additionally, combined with sender reputation

management, mail source verification and other

technologies, behavior-based filtering can effectively

improve the accuracy of detection and prevent false

alarms and misjudgments. Outlined below are several

key methods and their specific implementations.

4.2.1 User Behavior Monitoring

The behavior-based filtering method will firstly

determine whether the email is spam by monitoring

how user interacts with it. For example, the system

will record whether the user opens the email, clicks

on the link in it, or marks it as spam. Such behavior

data helps the system identify the types of mail that

users are not interested in or suspect that it is spam.

And if certain emails are often labelled as spam by a

large number of users, the system will raise the level

of suspicion of the sender's email, so as to treat emails

from the same source more strictly in future filtering

(Beckel, 2024).

4.2.2 Sender’s Reputation Management

Filtering methods based on the sender's reputation is

a core component of behavior filtering. The spam

filtering system will continuously monitor and

evaluate the sender's history, mail delivery behavior

and interaction with the recipients. And by using

techniques such as blacklist and whitelist, the filter

can make a preliminary judgment based on the

sender's reputation. The blacklist has recorded the

senders who are identified as sending spam, and their

emails will be blocked directly; while the whitelist

has listed reliable senders whose emails will be

filtered first. And besides, the behavior filtering

system will also refer to the sender's historical

delivery rate, opening rate, complaint rate and other

data. If emails from a particular sender are frequently

ignored or reported as spam, the system reduces the

sender’s reputation score, potentially affecting their

future delivery rates (Beckel, 2024; Blanchard, 2024).

4.2.3 Interaction Rate Analysis

The interaction rate is an important measurement

based on behavior filtering. Through analyzing the

opening rate, click rate, reply rate and other key

indicators of the email, the system would evaluate

whether the email meets the user's expectations. And

if the interaction rate of an email is significantly lower

than that of a normal email, and a large number of

users choose to ignore or delete the email, then the

email may be labelled as spam by the system. Emails

with high interaction rates usually come from trusted

senders, while emails with low interaction rates may

be spam, especially when other emails from the

sender also show similar low interaction

characteristics (Beckel, 2024).

4.2.4 Mail Source Verification Technology

(SPF and DKIM)

Email source verification is an important step to

prevent spam in the hybrid filtering method. And the

Sender Policy Framework (SPF) and Domain Key

Identification Mail (DKIM) are two commonly used

technologies. SPF confirms whether the sending

server of the mail has the right to send mail on behalf

of the domain to prevent the email address from being

forged. And DKIM verifies whether the mail is

tampered during transmission through digital

signature. Through these mail source verification

technologies, the system can determine whether the

mail comes from a legal sender, thus improving the

accuracy of filtering and reducing spam from fake

domain names. These technologies are widely used in

systems like SpamAssassin to verify mail through

SPF and DKIM records (Blanchard, 2024; Van Impe,

2016).

4.3 Machine Learning-Based Filtering

In modern spam filtering systems, machine learning

technologies play a crucial role. By analyzing and

classifying large amounts of email data, these

technologies could effectively distinguish between

spam and ham emails. Outlined below are several

commonly used machine learning methods and their

applications in spam filtering.

4.3.1 Naive Bayes

Naive Bayes is a probability-based classification

algorithm which is widely used in text classification

tasks. This model assumes that each word in an email

appears independently and predicts the category of

the email based on the conditional probability of these

DAML 2024 - International Conference on Data Analysis and Machine Learning

454

words. And Naive Bayes will learn from past emails

and calculate the possibility of each word appearing

in spam or ham emails. Thus, when a new email

arrives, the system would calculate the possibility of

it being spam based on the combination of words.The

classifier model uses a function that assigns the object

(such as an email) to a class ˆC

= C

from the existing

classes {C

, . . . , C

}, as described in Equation

(1) (Aragão et al., 2021).

𝐶



𝑎𝑟𝑔 max

∈,..,

𝑝𝐶



𝑝𝑥



| 𝐶











(1)

Fast computing speed and easy implementation are

the main strength of Naive Bayes, especially when

handling large amounts of text (Ghosh & Senthilrajan,

2023).

As shown in Figure 1, the Naive Bayes model

illustrates the class probabilities P(c) directly

depending on the features x

, x

, . . . , x

, and no

probability between the features is taken into account.

Figure 1: Naive Bayes Model Visualization of Class

Probabilities model (based on Aragão et al., 2021)

4.3.2 Support Vector Machines (SVM)

SVM distinguishes between spam and ham emails by

building a hyperplane. The fundamental goal of SVM

is to identify the best decision-making boundary to

keep distinct mail categories as separate as possible

(Ghosh & Senthilrajan, 2023). Equation (2) can be

used to describe the hyperplane.

H = VX + c (2)

where c is a constant and V is the vector (Gibson,

Issac, Zhang, & Jacob, 2020). And SVM is suitable

for processing high-dimensional data, such as word

vectors in emails, and can still perform well even with

few training samples (Ghosh & Senthilrajan, 2023).

In spam filtering, the strength of SVM is that it can

effectively handle emails with blurred boundaries and

minimizes the risk of misjudgment.

4.3.3 Random Forest

Random Forest is an integrated learning algorithm

that creates numerous decision trees and combines

their results to make classifications (Gibson et al.,

2020). And each decision tree makes binary decisions

about the characteristics in the content of emails to

narrow the scope of categories layer by layer. Having

many decision trees makes the random forest

algorithm have high robustness and high accuracy

compare to other machine learning algorithms. And

in spam filtering, the random selection of features in

random forests can reduce over-fitting and improve

adaptability to new spam (Ghosh & Senthilrajan,

2023).

4.3.4 Ensemble Learning

Ensemble learning methods can combine multiple

classifiers to improve the accuracy of spam detection.

The typical ensemble algorithms include Adaptive

Boosting (AdaBoost) and Bootstrap Aggregating

(Bagging). AdaBoost gradually trains multiple weak

classifiers by increasing the weight of misidentified

samples in previous classification and finally forms a

powerful classifier. In spam email classification,

Bagging is used to count the frequency of spam-

related words in the training dataset. It performs

multiple classifications based on bootstrap samples of

the training data. Finally, it will combine all the

results to make a final prediction (Ghosh &

Senthilrajan, 2023). Research has showed that

ensemble learning methods can significantly improve

spam filtering accuracy by combining the strengths of

multiple classifiers (Kumar, Sonowal, & Nishant,

2020).

4.4 Deep Learning-Based Filtering

Recently, deep learning models have made significant

progress in the field of spam filtering, surpassing the

traditional machine learning methods. Unlike

traditional models that rely on feature engineering,

deep learning models can automatically extract

complex features, especially when handling long text

sequences and complex spam behaviors. Outlined

below are several widely used deep learning models

and their specific applications in spam filtering.

4.4.1 Convolutional Neural Networks

(CNNs)

Although CNNs were initially performed well in

image recognition, they are also suitable for text

classification tasks. By treating email text as a one-

Spam Filtering in the Modern Era: A Review of Machine Learning, Deep Learning, and System Comparisons

455

dimensional data sequence, CNNs can extract local

text patterns through multiple convolutional layers.

For example, CNNs can capture common words,

phrases, or specific sentence structures which are

often found in spam emails. And with the convolution

operation and max pooling layers, CNNs could

identify representative features from the emails and

then use them to classify spam or ham emails (Roy,

Singh, & Banerjee, 2020). CNNs ’ high

computational efficiency allows for the rapid

processing of large-scale email data, particularly

useful for recognizing short text segments and fixed

patterns.

4.4.2 Recurrent Neural Networks (RNNs)

and Its Variants

RNNs perform well in natural language processing

tasks because of their ability to process serial data. In

spam filtering, RNNs can effectively capture the

timing dependency relationship in emails, helping to

understand the contextual meaning of the email

content. However, standard RNNs easily encounter

gradient disappearance problems when handling long

sequences, and this is the reason why Long Short-

Term Memory (LSTM) and Gated Recurrent Units

(GRU) are widely used to replace RNNs (Roy et al.,

2020). These variants can retain contextual

information over longer time periods and have a

better ability to capture the long text structure of the

email. For example, LSTMs can remember the key

phrases that appear throughout an email, helping

distinguish between emails that seem similar but have

different meanings. Besides, GRU could simplify

structure so that it can improve computing efficiency

while maintaining excellent performance, which

makes it suitable for application environments with

limited resources (Regina, Gopu, & Govindasamy,

2020).

4.4.3 Transformer-Based Models

Recently, Transformer-based models have shown

excellent performance in text processing tasks, and

BERT is a representative of it. BERT is pre-trained

on large datasets such as English Wikipedia,

containing 2.5 billion words, and BookCorpus, with

800 million words (Kaddoura, Alfandi, & Dahmani,

2020). Unlike traditional RNNs, Transformers can

use self-attention mechanisms to consider the

relationships between all words in a text at the same

time, thus capturing deeper semantic meaning in

emails (AbdulNabi & Yaseen, 2021). And BERT

further improves this process by using a bidirectional

encoder to understand the context of words, which

allows it to interpret word meanings more accurately

in different contexts. This makes it perform

outstanding in spam filtering, especially when dealing

with complex spam strategies and ambiguous words.

And additionally, through the pre-trained BERT

model, the spam filtering system can make use of a

large amount of knowledge from the external dataset

to improve the accuracy of recognition (Tida & Hsu,

2020).

4.4.4 Hybrid Models and Ensemble Learning

In practice, spam filtering systems often combine

multiple deep learning models to improve robustness

and accuracy. Combining CNNs with RNNs (such as

LSTM), Hybrid models can not only capture the local

features in emails but also retain context information

in the long sequences. Additionally, ensemble

learning methods are also widely applied in deep

learning models, using predictions from multiple

models to make a final decision. For example, multi-

layer classifiers can gradually improve detection

accuracy and reduce the occurrence of false alarms

and omissions (Ablel-Rheem, Ibrahim, Kasim,

Almazroi, & Ismail, 2020)

5 COMPARISON OF EXISTING

SPAM FILTERING SYSTEMS

Spam filtering algorithms are commonly assessed

using metrics such as accuracy, precision, recall, and

F1 score.

Accuracy: It is defined as the percentage of all

emails (true positives, false positives, true negatives,

and false negatives) that are correctly identified,

including both true positives (T

) and true negatives

). The formula for calculating accuracy is

represented in Equation (3) (Ablel-Rheem et al., 2020;

Adnan, Imam, Javed, & Murtza, 2024).

Accuracy =





 











 



 



)

Precision: Precision refers to the percentage of

emails that are successfully identified as spam (true

positives, T

) out of all emails that are projected to be

spam (true positives and false positives, T

+ F

)， as

calculated by Equation (4) (Ablel-Rheem et al., 2020;

Adnan et al., 2024).

Precision =









 



)

DAML 2024 - International Conference on Data Analysis and Machine Learning

456

Recall: Recall quantifies the percentage of real

spam emails that the model successfully classified，

calculated as shown in Equation (5). True positives

and false negatives (T

+ F

) are both taken into

account (Ablel-Rheem et al., 2020; Adnan et al.,

2024).

Recall =









 



(5)

F1-Score (F1): The F1-score is calculated as the

precision and recall harmonic means, as shown in

Equation (6) (Ablel-Rheem et al., 2020; Adnan et al.,

2024).

F1-Score = 2∗

 ∗ 

  

(6)

Table 1 presents performance metrics of five key

models across datasets, highlighting accuracy,

precision, recall, and F1 scores (Ghosh & Senthilrajan,

2023).

Table 1: Performance Metrics of Various Spam Filtering

Models (Ghosh & Senthilrajan, 2023).

Model Accuracy

(%)

Precision Recall F1

Score

Naive

Bayes

87.63 0.82 0.85 0.83

Support

Vector

Machines

95.10 0.93 0.94 0.93

Random

Forest

99.91 0.99 0.99 0.99

BERT 98.67 0.98 0.99 0.98

CNN 98.50 0.97 0.98 0.98

Random Forest and BERT consistently

demonstrate outstanding performance. Random

Forest achieves near-perfect accuracy in multiple

datasets, largely due to its ability to handle diverse

and complex data structures. On the other hand,

BERT's advantage lies in its contextual understanding

of email content, enabling it to analyze more intricate

relationships within the text. Naive Bayes, while

widely used for its simplicity and efficiency, tends to

perform less effectively compared to more advanced

models like SVM and CNN. The support vector

machines offer strong boundary definitions between

spam and legitimate emails, contributing to their high

recall rates. CNN, particularly useful in analyzing

data with spatial hierarchies such as email text, also

performs well but falls slightly behind transformer

models like BERT in terms of accuracy

(Ghosh &

Senthilrajan, 2023).

6 CHALLENGES AND FUTURE

TRENDS IN SPAM FILTERING

One of the critical challenges in spam filtering is the

issue of concept drift, which refers to the gradual

evolution of spam characteristics over time. As

spammers continually refine their techniques,

traditional models struggle to adapt to new spam

patterns. This results in the need for more dynamic,

adaptable spam detection systems.

The rise of Large Language Model (LLM)-

modified spam, where Large Language Models

(LLMs) are used to craft highly sophisticated emails

that bypass conventional filters, presents another

growing threat. Existing models are vulnerable to

these new types of spam that mimic legitimate

communication patterns with high accuracy.

To address these challenges, researchers are

exploring lifelong learning models like Error

Learning Continuous Adaptive Detection Process

(ELCADP), which are capable of continuously

updating their knowledge base and adapting to new

spam trends. This approach ensures that the model

can evolve alongside the changing landscape of spam

emails.

In the future, Artificial Intelligence (AI)-driven

automation and personalized filtering systems will

become key trends. These systems will rely on

individual user behavior and preferences to fine-tune

the filtering process, thereby minimizing false

positives and false negatives. Models like BERT and

transformers are expected to remain at the forefront

of these innovations, leveraging their ability to

process complex contextual information to improve

filtering accuracy.

7 CONCLUSION

Spam filtering technology has developed from a basic

rule-based system to a field dominated by advanced

machine learning and deep learning models. Models

such as Random Forest and BERT perform well in

spam detection, especially when dealing with

complex and diverse mail patterns. However, the

continuous evolution of spam policies, such as the

emergence of LLM-modified spam, continues to

challenge existing filters.

The future of spam filtering lies in building a

dynamic system that can adapt to the changing

characteristics of spam. Lifelong learning model and

AI-driven automation will play a key role in

maintaining the accuracy and reliability of the spam

Spam Filtering in the Modern Era: A Review of Machine Learning, Deep Learning, and System Comparisons

457

detection system. Looking to the future, integrating

these advanced technologies is crucial to ensure that

users have a safer and more efficient email

environment.

REFERENCES

Aragão, M. V. C., Ferreira, I. C., Oliveira, E. M., Kuehne,

B. T., Moreira, E. M., & Carpinteiro, O. A. S., 2021. A

study and evaluation of classifiers for anti-spam

systems. IEEE Access, 9, pp. 157482-157496.

AbdulNabi, I., & Yaseen, Q., 2021. Spam email detection

using deep learning techniques. Procedia Computer

Science, 184, 853-858.

Ablel-Rheem, D. M., Ibrahim, A. O., Kasim, S., Almazroi,

A. A., & Ismail, M. A., 2020. Hybrid feature selection

and ensemble learning method for spam email

classification. International Journal of Advanced

Trends in Computer Science and Engineering, 9(1.4),

217-223.

Adnan, M., Imam, M. O., Javed, M. F., & Murtza, I., 2024.

Improving spam email classification accuracy using

ensemble techniques: A stacking approach.

International Journal of Information Security, 23, 505–

517.

Beaman, C., & Isah, H., 2021. Anomaly detection in emails

using machine learning and header information.

Canadian Institute for Cybersecurity, University of

New Brunswick.

Beckel, Z., 2024. How email spam filters work:

Understanding your email filter system. Qudni.

Blanchard, T., 2024. An in-depth guide to how email spam

filter works. Sterling Technology Solutions.

Chu, K.-T., Hsu, H.-T., Sheu, J.-J., Yang, W.-P., & Lee, C.-

C., 2020. Effective spam filter based on a hybrid

method of header checking and content parsing. IET

Networks, 9(6), 338-347.

Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K., 2019.

BERT: Pre-training of deep bidirectional transformers

for language understanding. arXiv.

Elliott, A., 2023. SpamAssassin scores explained: What

email marketers need to know. MailerCheck.

Feng, W., Sun, J., Zhang, L., Cao, C., & Yang, Q., 2016. A

support vector machine based naive Bayes algorithm

for spam filtering. 2016 IEEE 35th International

Performance Computing and Communications

Conference (IPCCC), 1-8.

Ghosh, A., & Senthilrajan, A., 2023. Comparison of

machine learning techniques for spam detection.

Multimedia Tools and Applications.

Gibson, S., Issac, B., Zhang, L., & Jacob, S. M., 2020.

Detecting spam email with machine learning optimized

with bio-inspired metaheuristic algorithms. IEEE

Access, 8, 187914-187932.

Junnarkar, A., Adhikari, S., Fagania, J., Chimurkar, P., &

Karia, D., 2021. E-Mail spam classification via

machine learning and natural language processing.

2021 Third International Conference on Intelligent

Communication Technologies and Virtual Mobile

Networks (ICICV), 693-699.

Kaddoura, S., Alfandi, O., & Dahmani, N., 2020. A spam

email detection mechanism for English language text

emails using deep learning approach. 2020 IEEE 29th

International Conference on Enabling Technologies:

Infrastructure for Collaborative Enterprises (WETICE),

193-198.

Kumar, N., Sonowal, S., & Nishant., 2020. Email spam

detection using machine learning algorithms. 2020

Second International Conference on Inventive Research

in Computing Applications (ICIRCA), 108-113.

Kaddoura, S., Alfandi, O., & Dahmani, N., 2020. A spam

email detection mechanism for English language text

emails using deep learning approach. 2020 IEEE 29th

International Conference on Enabling Technologies:

Infrastructure for Collaborative Enterprises (WETICE),

193-198.

Mohammad, R. M. A., 2020. A lifelong spam emails

classification model. Applied Computing and

Informatics, 20(1/2), 35-54.

Roumeliotis, K. I., Tselikas, N. D., & Nasiopoulos, D. K.,

2024. Next-generation spam filtering: Comparative

fine-tuning of LLMs, NLPs, and CNN models for email

spam classification. Electronics, 13(2034).

Roy, P. K., Singh, J. P., & Banerjee, S., 2020. Deep learning

to filter SMS spam. Future Generation Computer

Systems, 102, 524-533.

Regina, K., Gopu, A., & Govindasamy, V., 2020. Spam

detection using recurrent neural networks. International

Journal for Research in Engineering Application &

Management, 6(1), 313-318.

Saidani, N., Adi, K., & Allili, M. S., 2020. A semantic-

based classification approach for an enhanced spam

detection. Computers & Security, 94, 101716.

Si, S., Wu, Y., Tang, L., Zhang, Y., & Wosik, J., 2024.

Evaluating the performance of ChatGPT for spam email

detection. arXiv.

Tida, V. S., & Hsu, S., 2020. Universal spam detection

using transfer learning of BERT model. Center of

Advanced Computer Studies, University of Louisiana

at Lafayette.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I., 2017.

Attention is all you need. In 31st Conference on Neural

Information Processing Systems (NIPS 2017), Long

Beach, CA, USA.

Van Impe, K., 2016. Understanding the SPF and DKIM

spam filtering mechanisms. Security Intelligence.

Zych, S., 2020. Patterns are power: A beginner’s guide to

spam detection. Medium.

DAML 2024 - International Conference on Data Analysis and Machine Learning

458