FakeIDNet: Identifying Suspicious Profiles with Artificial Neural Networks
ABSTRACT
• In this paper, we use machine learning, namely an artificial neural network to determine
what are the chances that Facebook friend request is authentic or not. We also outline
the classes and libraries involved. Furthermore, we discuss the sigmoid function and
how the weights are determined and used. Finally, we consider the parameters of the
social network page which are utmost important in the provided solution.
Introduction:
• In today's digital age, the ever-increasing dependency on computer technology has left
the average citizen vulnerable to crimes such as data breaches and possible identity
theft. These attacks can occur without notice and often without notification to the
victims of a data breach. At this time, there is little incentive for social networks to
improve their data security. These breaches often target social media networks such as
Facebook and Twitter. They can also target banks and other financial institutions.
Problem Identification:
• The fake profile's contents typically have links that lead to an external website where
the damage happens. An unaware curious user clicking the bad link will damage their
computer. The cost can be as simple as catching a virus to as bad as installing a rootkit
turning the computer into a zombie. While Facebook has a rigorous screening to keep
these fake accounts out, it only takes one fake profile to damage the computers of many.
Existing System:
• The algorithm uses a seed selection method combined with early terminated random
walks to propagate trust [5]. Its computational cost is measured in O (nlogn). Profiles
are ranked according to the number of interactions, tags, wall posts, and friends over
time. Profiles that have a high rank are considered to be real with fake profiles having
a low rank in the system. Unfortunately, this technique was found to be mostly
unreliable because it failed to take into account the possibility that real profiles can be
ranked low and fake profiles can be ranked high.
Limitations of the Existing System:
Although Facebook and similar platforms apply strict screening techniques to block fake
profiles, the current systems still face several critical limitations:
1. Post-Creation Detection Delay
Fake profiles are often detected only after they start spreading malicious links. By
then, many users may already be affected.
2. Link-Based Attacks Are Not Prevented Early
The existing system does not effectively stop fake profiles before they share harmful
URLs.
Once a user clicks the malicious link, the damage (virus, malware, rootkit) has already
occurred.
3. One Fake Profile Can Harm Many Users
Even with screening, a single fake account is enough to spread malware to hundreds
of users, making the system highly risky.
4. Lack of Behavioral Intelligence
Current systems mainly focus on profile verification and content filtering but fail to
analyze:
• Posting behavior
• Friend request patterns
• Message repetition
• Suspicious activity frequency
5. Reactive Instead of Proactive
The system reacts after harm occurs, rather than predicting and preventing fake
profiles at the time of friend requests.
6. High False Negatives
Some fake profiles successfully bypass checks and remain active long enough to
cause serious damage.
Technologies Used:
The existing system relies on traditional, rule-based and graph-based techniques to
identify fake profiles.
Technology Description
Graph-Based Algorithms Model social networks as graphs of users and connections
Trust Propagation Methods Spread trust values across connected profiles
Ranking Algorithms Rank profiles using interaction counts
Statistical Analysis Count friends, posts, likes, tags
Rule-Based Classification Uses predefined thresholds
Traditional Databases Store profile and interaction data
Batch Processing Systems Periodic profile ranking
Proposed System:
• We use machine learning, namely an artificial neural network to determine what are the
chances that a friend request is authentic or not. Each equation at each neuron (node) is
put through a Sigmoid function to keep the results between the interval of 0.0 and 1.0.
At the output end, this could easily be multiplied by 100 to give us the possible
percentage that it is a malicious request. Our solution would be only one deep neural
network, meaning it only has a single hidden layer.
Advantages:
1. High Detection Accuracy
Unlike rule-based systems, the ANN learns patterns from real data, allowing it to
correctly classify both genuine and fake profiles even when behavior is complex or
hidden.
2. Probability-Based Output
The sigmoid function produces a value between 0 and 1, which can be converted into
a percentage of authenticity.
This gives a clear confidence level instead of just a yes/no result.
3. Early Fake Profile Detection
The system can detect malicious profiles at the time of the friend request, before the
user interacts or clicks harmful links.
4. Handles Complex Behavior Patterns
Even with a single hidden layer, the neural network can model non-linear
relationships between profile features such as activity, friends, posts, and links.
5. Reduced False Positives and Negatives
The ANN does not rely on fixed thresholds. It adapts based on training data, reducing:
• False positives (real users marked fake)
• False negatives (fake users marked real)
6. Adaptive and Self-Learning
When new data is added, the network can be retrained, allowing the system to evolve
with new attack strategies.
7. Efficient and Scalable
A single hidden-layer ANN has lower computational cost than deep multi-layer
networks, making it fast and suitable for large-scale social networks.
8. Improved User Safety
By stopping fake requests early, the system protects users from:
• Malware
• Phishing links
• Account hijacking
Technologies Used:
The proposed system uses machine learning and Artificial Neural Networks.
Technology Description
Python 3.7 Core programming language
Artificial Neural Network (ANN) Learns patterns to classify fake profiles
TensorFlow / Keras ANN model creation and training
NumPy & Pandas Feature extraction and data processing
Sigmoid Activation Function Generates probability (0–1) output
Scikit-learn Model evaluation
Django Framework Web-based interface
SOFTWARE REQUIREMENTS:
Programming Language: Python 3.7.0
Framework: Django
Libraries: NumPy, Pandas, Scikit-learn, TensorFlow/Keras
Database: SQL
Conclusion:
• We use machine learning, namely an artificial neural network to determine what are the
chances that a friend request is authentic are or not. Each equation at each neuron (node)
is put through a Sigmoid function. We use a training data set by Facebook or other
social networks. This would allow the presented deep learning algorithm to learn the
patterns of bot behavior by back propagation, minimizing the final cost function and
adjusting each neuron's weight and bias.
• In this paper, we outline the classes and libraries involved. We also discuss the sigmoid
function and how are the weights determined and used. We also consider the parameters
of the social network page which are the most important to our solution.