Comparative Analysis of Machine Learning

Classifiers for Fake Account Detection

Yashwant Dongre Pranit Dalavi Adarsh Dawkhare

Computer Dept Computer Dept Computer Dept
VIIT, Pune VIIT, Pune VIIT, Pune
yashwant.dongre@viit.ac.in pranit.22010875@viit.ac.in adarsh.22011017@viit.ac.in

Suraj Karle improves the quality of user interactions by reducing spam,

Computer Dept fake engagement, and irrelevant content. This, in turn,
VIIT, Pune enhances the overall user experience on social media platforms.
suraj.22010680@viit.ac.in To verify the content is fake or genuine various ML approaches
such as KNN, SVM, Logistic Regression are used.[1] In the
proposed system ,SENAD method and CredNN model is used
to recognize fraud news, fraudulent accounts, and fake images
on online platforms.. These models analyze social and
Abstract— Authenticity is vital for maintaining trust on social
psychological aspects, image authenticity, and user account
media platforms. Fake accounts can spread misinformation,
engage in malicious activities, or impersonate individuals,
engagement patterns to identify fake news. The model uses
eroding trust and credibility within the online community. features such as account age, follower-to-following ratio, and
Financial scams, identity theft, phishing, and other fraudulent influential power of nodes in the network to determine the
activities frequently result in the creation of malicious authenticity score of users and posts[2]. The proposed approach
accounts. Detecting and eliminating these accounts helps makes use of a Chrome extension to evaluate features and apply
prevent users from falling victim to such activities. The system Mahine_Learning classifiers to divide the dataset into authentic
addresses fake account detection on social media through and fraudulent accounts. Petri net-based solution is employed
machine learning, enhancing online security by identifying to recognize the source of fake content. The framework also
fake profiles. Utilizing advanced machine learning algorithms
involves the analysis of various Twitter characteristics, such as
the system distinguishes patterns and anomalies, effectively
detecting fake social media accounts. This system uses
tweets, to distinguish between real and false accounts and
innovative features and deep learning techniques to detect ultimately contributing to the security and trustworthiness of
malicious accounts. This system excels with an accuracy rate social media platforms [3]. The DeepFake is a term used by the
exceeding 95%, minimizing false positives. It attains a recall community to describe face picture and video manipulation
rate of 90%, effectively identifying the fake accounts. The using AI or deep-learning. The suggested method makes
system demonstrates scalability. It contributes to a reliable and advantage of DeepFake technology, which uses deep
efficient solution for fake account detection in dynamic online generativetechniques to create and modify physical
environments. This system combats rising cyber threats, appearance. [4].In the proposed system different methods are
securing online spaces, fostering trust, and maintaining the
used to detect spammerson Twitter .SVM, Random Forest, and
integrity of social media interactions. In this way the system
provides relevance to the real world. Multilayer Perceptron (MLP) classifiers has been used for
spam recognition. PCA with tunedK- means algorithm has been
used to find favorableresults with the number of clusters
Keywords—LSTM , Recall , Accuracy, Fake , SVM,ANN examined and to find detection of malicious accounts. A DBMS
stores the data that was taken from the Twitter REST API for
future analysis.[5].
I. INTRODUCTION

II. LITERATURE REVIEW

Fake accounts can be used for various malicious activities,
including manipulating trends, gaming algorithms, or
influencing public opinion. Detecting and removing fake Classification of Twitter scammer detection methods is
accounts preserves the integrity of the platform and ensures a proposed in the system. The researchers categorized the
fair and authentic user experience. Eliminating fake accounts approaches into different categories, such as fake content
identification , URL spam recognition , detecting spam in embedding techniques like word2vec .To illustrate the MVAN
trending subjects and malicious account recognition. It contains model's efficacy and resilience in early fake news
the comparison of various Twitter Spam detection features identification, the study contrasts it with several cutting-edge
,broken down into 5 categories : time, structure, graph, user and models, such as SVM-BOW, BiLSTM, and TextCNN.
content. Each class utilizes models such as regression prediction, Accuracy of MVAN method is significantly improved than
ML, Naïve Bayes classifier ,hybrid techniques, clustering other models, with improvements ranging from 2.03% to 25%
algorithm and language model. The proposed method using on different datasets.[11].
machine learning classifiers produced the best accuracy results,
The System involves a grouping of Natural Language
with random forest operating on a collection of 19 features-set.
Processing (NLP) and Reinforcement Learning (RL) for
[6] . The System presents a novel based method for quickly
detection and prohibition of fake accounts. Block-chain and
identifying newly registered Twitter accounts, tracking their
NLP uses ML techniques to detect andcalculate fake customer
activity, and assessing the safeguards. Twitter-related studies
accounts. CNNs for image classification and text classification
suggest gathering tweets using queries connected to specific
have been used. The system involves existing investigation
areas or using an API to sample random tweets in real time.
associated to OSN security threats, earlymalicious accounts
Snowflake IDs are used for collecting account information for
and Sybil recognition methods . Three main categories used to
analysis. In a shared environment, it manages the establishment
classify social network bots are network structure, content-
of time- ordered ID for customer accounts and tweets[7]. The
based defense and hybrid techniques. Collection of profiles that
EGSLA method is used to recognize false accounts from huge
are accessible on the network are used in the LinkedIn dataset
amount of information. EGSLA technique consists of four
to recognize fake profiles. A multi-stagelevel classification
modules which are data grouping, feature extraction,
process is used by Sybil Frame to find Sybils on Facebook and
organization and decision making. It uses machine learning
Twitter. [13]. In the proposed system the algorithm uses
algorithms and parameterized approaches to detect fake users
Sentimental_analysis on data which is gathered on a specific
and spam content, with a focuson improving accuracy and
topic of interest of users .In order to identify spam on Twitter,
reducing the spam impact . The information is composed with
this algorithm calculates each user's sentiment score using
the help of Python web-scraping structure. Feature extraction
content-based criteria. In order to determine how effective the
exhibits formulation for all the features used to identify false
system is in identifying fakeaccounts, the suggested method
customers. A useful method for accessing records on
compares its own rule-based algorithm for bot detection with a
categorization issues is graph-based learning. The EGSLA
few others, including MLP, decision trees, and random forests
gained an accuracy of 90.3%, a true F1- score of 88.2%, and a
on real time data .The fake prediction algorithm achieved an
false F1-Score of 89.7%, outperforming other ML Classifiers
accuracy of 97.67%, which was higher than other machine
such as KNN, SVM [8].The proposed system consists of the
learning classifiers such as decision tree (81.4%), MLP (78.3%),
DeepProfile dynamic CNN model for recognising fake accounts
and random forest (78.34%) [14]. The System involved data
on social media sites. This model isexperienced with
mining from Twitterusing the twitter API and storing the data in
comprehensive study to get state- ofthe-art outcomes, especially
a non- relational database. It consists of different features such
for identifying false accounts. Models Similar to the Learning
as the corpus, noting characteristics such as few friends and
Model , CNN design and Dynamic CNN are used to deal with
followers, sharedprofile descriptions, and the presence of URLs
false account issues. It involves utilizing different methods to
in some descriptions .Cross-validation and re-sampling are used
recognize and categorize irregularities in the OSN environment.
to perform supervised ML systems to recognize
This includes the use of behavioral_analysis and learning
accountsthathave been identified as fake in the dataset. Three
methods, The DeepProfile approach demonstrates higher
ML models such as random forest, boosting, and SVM were
precision of 94.00%. The classification task performance of
applied. [15].
CNN can be enhanced usingthis Model.[9].
The primary limitations observed in the system relate to
The System consists of methods for recognising malicious
models not operating in real-time, leading to delays in
accounts in OSNs using a LR classifier with a
identifying and mitigating fake accounts. To address this, there
particle_swarm_optimization (PSO) model. The experimental
is a need to develop models and techniques capable of real-time
results show an improvement in the accuracy of the architecture
detection to promptly respond to emerging threats. Another
from 93.7% to 96.5. The Naive Bayes classifier achieved an
challenge identified is the use of imbalanced datasets, where
accuracy of 84.6%, the Decision Tree classifier achieved an
the number of fake accounts is significantly smaller than
accuracy of 88.03%, the Logistic Regression classifier
genuine ones, posing difficulties in training robust models.
achieved an accuracy of 93.3%, and the Logistic Regression
with PSO gained the highest accuracy of 96.2%. [10]. The
System involves the use of DL techniques for false news
identification on social media, specifically Twitter. The Twitter III. METHODOLOGY
propagation structure of tweets is analyzed using the MVAN The Model has been implemented using 6 algorithms . For
Model. The model records the relationships between words and this system it uses 2 datasets “real_users” and “fake_users”.
generates the original tweet representation using word
“real_users” contains information of real accounts of users and cleaning, transforming, and organizing raw data into format
“fake_users” contains information of fake accounts of users. that is appropriate for investigation and modeling. Data needs
The dataset used to detect fake accounts from social media to be processed before extracting features from it. For the pre-
networks was sourced from Kaggle and comprises a total of processing of data, first the punctuations were removed,
4000 rows ,2000 for real users and 2000 for fake users. lowercasing of words was done, removal of stopwords, and
Subsequent to the dataset loading phase, data preprocessing has lemmatization was performed, after which tokenization and
been executed. Preprocessing data is a crucial stage in the padding of sentences were performed.
pipeline for machine learning and data analysis. It requires

B. SYSTEM DESIGN AND IMPLEMENTATION

performance of the model, common classification metrics
are used. These metrics include the ROC curve, accuracy,
precision, recall, and F1-score. Fig. 1 illustrates how the
algorithm operates.

A. DATA ACQUISITION AND PRE-PROCESSING

The dataset has been taken from Kaggle and is used for
detecting fake accounts. In all, there are 4000 rows in it.
Real users count 2000, while fake users count 2000. ID,
name, Statuses_Count, followers_count, friends_count,
and favorites_count make up the dataset.
An essential phase in the pipeline for data analysis and
machine learning is data preprocessing. Several
techniques are used in the process, including lowercasing,
stopword removal, stemming, tokenization, and padding.
Lowercasing, a common preprocessing technique,
involves converting all text data to lowercase. This allows
models to focus on the semantic content of the text,
mitigating the impact of variations in letter case. Stopword
removal entails eliminating common words, known as
stop words, from the text data. This step concentrates on
retaining content-rich words, enhancing the meaningful
information.Tokenization is a key preprocessing technique
that breaks down text into individual words or tokens,
aiding in the model's understanding of textual data.
The LSTM neural network processes textual descriptions
of user accounts by sequentially analyzing and learning
complex dependencies in the data. Its ability to capture long-
range dependencies helps in discerning subtle linguistic
patterns indicative of fake accounts. By embedding the text
and utilizing the LSTM layer, the model gains contextual
understanding, allowing it to recognize intricate linguistic
structures associated with fake profiles. This analysis
improves the systems capability to distinguish between
authentic and malevolent accounts by utilizing the sequential
Fig.1. Block Diagram of Architecture information encoded in the textual data. Traditional RNNs
have a drawbackof capturing the long- term dependencies
because as thenetwork processes new items, knowledge about
After that , The data was split into training and testing sets
earlier partsin the sequence fades away. This problem is
using the train test split method. After Splitting data, define
known as the gradient vanishing problem. In order to
the Models and compile it and test the model on training
overcome this problem, LSTM, which is Long Short-Term
data andplot the results .The User Account details are
Memory was designed.
processed usinga total 6 ML Models . To calculate the
 The LSTM process begins with loading and shuffling the patterns that are valuable for detecting fake accounts
dataset to ensure randomness. Text preprocessing steps follow, based on user descriptions
including converting text to lowercase, removing unnecessary
words, and tokenizing Numerical features are extracted, labels
are assigned, and data is split into training and testing sets.
LSTM model is used to organize users as either real or fake
based on their Twitter descriptions and other numerical Algorithm – Fake Account Detection Model using LSTM
features. The model first embeds the text descriptions into a
vector representation using an embedding layer. Then, it feeds
the embedded vectors into an LSTM layer, which learns to Input : User Account Details
capture long-term dependencies in the text. Finally, it
Output : Detecting Real and Fake Accounts
combines the output of the LSTM layer with the numerical
features and passes them through a dense layer to produce the 1 : Loading Dataset
final classification prediction. 2 : Pre-processing the Input text
Equations (1), (2) and (3) describes the gates present in the 3 : for every sentence in training sentences do
LSTM layer, where 𝑓𝑡, 𝑖𝑡 and 𝑜𝑡 represents the forget, input 4 : Remove punctuations
and output gates respectively.
5 : Perform Lowercasing of words
6 : Remove Stopwords
𝑓𝑡 = 𝜎(𝜔𝑓ℎ𝑡−1 + 𝑈𝑓𝑥𝑡 + 𝑏𝑓) - (1) Where, σ 7 : Perform tokenization and padding.
is sigmoid activation function 𝜔𝑓 is forget gate 8 : end for
weight matrix 9 : Applying LSTM Model
𝑏𝑓 is forget gate bias vector 10 : Compile, train, and fit the model

Equation 1 represents the Forget gate . Forget gate helps the 11 : Test the model and calculate the
LSTM model focus on relevant and informative featuresin the performanceparameters
user's description, discarding irrelevant and potentially 12 : Return whether an account is Real or Fake
misleading information.
𝑖𝑡 =𝜎(𝑤𝑖ℎ𝑡−1 + 𝑈𝑖𝑥𝑡 + 𝑏𝑖) - (2)
Where,
The Artificial Neural Network (ANN) learns patterns
𝑤𝑖 is input gate weight matrix and relationships within user features, enabling it to
𝑏𝑖 is input gate bias vector discern distinguishing characteristics of fake accounts.
By iteratively adjusting weights during training, the
Equation 2 represents the Input gate . Input gate decides model captures complex representations associated with
how much new data should be added by taking into deceptive behavior. The final trained ANN utilizes these
consideration the current input and previous state’s output. learned patterns to makepredictions, effectively
identifying potential fake accounts based on the input
feature patterns it has internalized. The Decision Tree
𝑜𝑡 = (𝜎𝜔0 ℎ𝑡−1 + 𝑈0𝑥𝑡 + 𝑏0 ) - (3) model aids in detecting fake accounts by recursively
Where, partitioning the feature space based on attribute values.
It makes binary decisions at each node to maximize
𝑤𝑜 is output gate weight matrix information gain or minimize impurity. During training,
𝑏𝑜 is output gate bias vector the model learns to distinguish between genuine and
fake accounts by identifying crucial feature thresholds.
Equation 3 represents Output gate . The output gate In the evaluation phase, the Decision Tree efficiently
generate a meaningful output in the form of a probability classifiesinstances by traversing the learned tree
value between 0 and 1. A value nearer to 1 specifies that structure.
the network is highly confident that the account is real, The SVM is a ML algorithm works by mapping the
while avalue closer to 0 specifies that the network is input features (such as user attributes) and finding a
highly confident that the account is fake. hyperplane that best separates the two modules (fake and
LSTM structure represented by these equations genuine accounts).The SVM optimization process
provides the model with the ability to analyze sequential involves maximizing the margin between the classes
data, effectively capturing dependencies and discerning while penalizing misclassifications. The 'C' and 'gamma'
parameters are tunedthrough a grid search to optimize
 the SVM's performance. The resulting SVM model Fig 2. Shows accuracy performance of model. The
classifies new instances by assigning them to one of the accuracy scores indicate the performance of different models
two classes based on their position relative to the in classifying fake accounts. Decision Tree and ANN exhibit
hyperplane in the feature space. The Random Forest the highest accuracies at 98.75% and 98.76%, respectively,
algorithm is an ensemble learning techniquethat builds showcasing their proficiency in distinguishing between
multiple decision trees and combines their predictions genuine and fake accounts. Random Forest follows closely
for improved accuracy. Every tree is trained on a random with an accuracy of 94.5%, indicating robust ensemble
subset of dataset, reducing overfitting and enhancing learning. XGBoost and SVM also demonstrate strong
generalization. The finalprediction is determined by a performance at 94.15% and 92.91%, respectively.
majority vote from the individual trees. This approach is
effective for detecting fake accounts as it can capture
complex patterns and relationshipswithin diverse user
attribute data, offering a robust solution for classification
tasks.
XGBoost, aids in fake account detection by
constructing arobust ensemble of decision trees that
collectively model complex relationships within the
dataset. It excels at handling imbalanced classes and
prioritizes correct classification of minority class
instances, which is crucial for identifying fakeaccounts.
The algorithm's optimization objective, which combines
both loss minimization and regularization terms,
enhances its ability to generalize well on unseen data,
contributing to its effectiveness in discerning genuine
and fake accounts.
Fig 3. Precision Score of different Algorithms

IV. RESULTS AND DISCUSSIONS

The Fig.3 shows the precision performance parameters by
Assessing a model's efficacy through metrics like F1 model for fake account detection. The precision scores
score, accuracy, precision, and recall offers a thorough picture indicate the amount of true positive calculations among all
of how successful it is. Accuracy measures the overall instances predicted as positive by each model. SVM and
correctness of predictions, precision quantifies the model's Decision Tree achieve perfect precision which is 1.0000,
capability to avoid false positives, recall assesses its capability signifying that all instances predicted as positive are indeed
true positives. Random Forest closely follows with an
to capture true positives, and F1 score strikes a balance between exceptionally high precision which is 0.9963. XGB classifier
precision and recall. Utilizing these metrics collectively ,ANN and LSTM is having precision values of 0.9852, 0.9769
ensures a exact evaluation, especially in scenarios with and 0.7455 respectively.
imbalanced class distributions, such as detecting fake accounts, Fig 4. Recall Score of different Algorithms
where false positives and false negatives carry distinct
significance. Fig 4. shows the recall performance parameter for fake

Fig.2.Accuracy Score of different Algorithms account detection. A classification model with a recall metric
can identify every related case of a given class. Greater The left-hand plot shows the k-means clusters after the
recall values show how well a model captures a significant algorithm has converged. The data points are colored
fraction of true positive cases in comparison to the real according to their assigned cluster. The right-hand plot
positives in the dataset. In the presented results, ANN shows the original data before clustering.
(Artificial Neural Network) attains the highest recall which
The k-means algorithm has successfully identified two
0.9961, reflecting its efficacy in identifying genuine
clusters of data points. The first cluster (shown in blue)
accounts among the total genuine instances. XGB Classifier
contains data points that are for the real accounts. The second
, SVM, Random Forest , LSTM and Decision Tree has
cluster (shown in red) contains data points that are for the fake
Recall values of 0.9020 , 0.8649 , 0.8986, 0.6753 and 0.9725
accounts. The k-means algorithm can be used to detect fake
respectively.
accounts by identifying data points that are anomalous or
outliers.

Fig 5. F1-Score of different Algorithms

Fig 7.ROC curve of algorithms

Fig 5. shows the F1-score performance parameter . ANN The area under the ROC curve (AUC) is a single metric
and Decision Tree exhibits a strong F1 score of 0.9864, that used to summarize the performance of classifier. A
highlighting its precision and recall balance. Random Forest higher AUC shows better performance. In Fig 5, the
follows closely with a score of 0.9449, indicating effective Random Forest classifier has highest AUC (0.87), followed
performance in correctly identifying both classes. XGB by the XGBoost classifier (0.86), the SVM classifier (0.86),
classifier, SVM and LSTM has f1-scores values of 0.9418, and the Decision Tree classifier (0.78).
0.9275 and 0.7087 respectively. III. CONCLUSION
Successfully detected the massive problem of malicious
accounts on online social platforms using advanced
machine learning algorithms . The proposed solution
maintains the authenticity of social media interactions and
enhances online environments due to its Accuracy ,
precision, Recall, adaptability, and scalability. The
existing system has some novelties such as, by focusing on
the behavioral patterns of user interactions, the system
provides an innovative approach and provides an
advanced understanding that enhances the detection of
fake accounts. Using advanced deep learning and Machine
learning algorithms, the system exceeds traditional
methods and accurately detects fake profiles. Innovative
Fig 6. Scatter plot using K-means Clustering optimization techniques reduce false positives and false
negatives while maintaining a special balance between
The Fig.6 displays the outcomes of applying k-means precision and recall which increases the overall efficiency
procedure to a dataset of fake account detection . The k- of detecting fake accounts. According to above
means algorithm is unsupervised learning method that observation ANN has maximum Accuracy of 98.76%,
clusters data points into k clusters based on their similarity.
 SVM and Decision Tree has maximum precision value of
1.00, ANN has maximum Recall value of 0.9961 and ANN
has maximum F1-Score value of 0.9864.The limitations of
the existingsystem consists of, the lack of diverse training
data mightaffect the systems performance by reducing the
ability todetect fake accounts.

REFERENCES
[1] Sahoo, Somya Ranjan, and Brij B. Gupta. "Multiple features based
approach for automatic fake news detection on social networks using deep
learning." Applied Soft Computing 100 (2021): 106983.
[2] Uppada, Santosh Kumar, K. Manasa, B. Vidhathri, R. Harini, and B.
Sivaselvan. "Novel approaches to fake news and fake account detection in
OSNs: user social engagement and visual content centric model." Social
Network Analysis and Mining 12, no. 1 (2022): 52.
[3] Sahoo, Somya Ranjan, and Brij B. Gupta. "Hybrid approach for detection
of malicious profiles in twitter." Computers & Electrical Engineering 76
(2019): 65-81..
[4] Juefei-Xu, Felix, Run Wang, Yihao Huang, Qing Guo, Lei Ma, and Yang
Liu. "Countering malicious deepfakes: Survey, battleground, and horizon."
International journal of computer vision 130, no. 7 (2022): 1678-1734.
[5] Adewole, Kayode Sakariyah, Tao Han, Wanqing Wu, Houbing Song, and
Arun Kumar Sangaiah. "Twitter spam account detection based on
clustering and classification methods." The Journal of Supercomputing 76
(2020): 4802-4837.
[6] Masood, Faiza, Ahmad Almogren, Assad Abbas, Hasan Ali Khattak, Ikram
Ud Din, Mohsen Guizani, and Mansour Zuair. "Spammer detection and
fake user identification on social networks." IEEE Access 7 (2019): 68140-
68152.
[7] Cola, Guglielmo, Michele Mazza, and Maurizio Tesconi. "Twitter
Newcomers: Uncovering the Behavior and Fate of New Accounts through
Early Detection and Monitoring." IEEE Access (2023).
[8] Balaanand, Muthu, S. Karthik,Gunasekaran Manogaran, and C. B.
Sivaparthipan. "An enhanced graph- based semi-supervised learning
algorithm to detect fake users on Twitter."The Journal of Supercomputing
75 (2019): 6085-6105.
[9] Wanda, Putra, and Huang Jin Jie. "DeepProfile: Finding fake profile in
online social network using dynamic CNN." Journal of Information
Security and Applications 52 (2020): 102465.
[10] Bharti, Kusum Kumari, and Shivanjali Pandey. "Fake account detection in
twitter using logistic regression with particle swarm optimization." Soft
Computing 25, no. 16 (2021): 11333-11345.
[11] Ni, Shiwen, Jiawen Li, and Hung-Yu Kao. "MVAN: Multi-view attention
networks for fake news detection on social media." IEEE Access 9 (2021):
106907-106917.
[12] Shahbazi, Zeinab, and Yung-Cheol Byun. "Fake media detection based on
natural language processing and blockchain approaches." IEEE Access 9
(2021): 128442-128453.
[13] Ramalingam, Devakunchari, and Valliyammai Chinnaiah. "Fake profile
detection techniques in large-scale online social networks: A
comprehensive review." Computers & Electrical Engineering 65 (2018):
165-177.
[14] Monica, C., and N. Nagarathna. "Detection of fake tweets using sentiment
analysis." SN Computer Science 1 (2020): 1-7.
[15] Van Der Walt, Estée, and Jan Eloff. "Using machine learning to detect fake
identities: bots vs humans." IEEE access 6 (2018): 6540-6549.