Online Fake Logo Detection System

Vivek Tanniru (  [email protected] )

Auburn university at Montgomery Montgomery
Tathagata Bhattacharya
Auburn university at Montgomery Montgomery

Research Article

Keywords:

Posted Date: January 20th, 2023

DOI: https://doi.org/10.21203/rs.3.rs-2492597/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Additional Declarations: No competing interests reported.

 Online Fake Logo Detection System

Vivek Tanniru Dr.Tathagata Bhattacharya,Asst.Professsor

Department of Computer Science Department of Computer Science
Auburn university at Montgomery Auburn University at Montgomery
Montgomery, Alabama,USA Montgomery,Alabama,USA
[email protected] [email protected]

Abstract—With the increasing prevalence of online fraud and A. Origin of Logo’s

the use of fake logos to deceive consumers, there is a need
for effective methods to detect and prevent the use of fake We’ve all heard that a logo can be a very powerful asset for
logos on the internet. In this paper, we propose a method for a brand. In fact, many businesses are recognized faster by their
detecting fake logos using machine learning techniques. Our logo than by their name. Logos abound, and well-known ones
approach involves extracting features from the logos and training are easily recognized by nearly every consumer. But where did
a classifier to distinguish between real and fake logos. We evaluate
the performance of our method on a dataset of real and fake logos
the logo come from? How has it changed over time? The logo’s
and demonstrate its effectiveness in detecting fake logos with high origins can be traced back to the Ancient Egyptians. They
accuracy. Every day, hundreds of domain names, websites and used hieroglyphics to brand and identify their possessions until
logos are being cloned by cyber criminals who want to gain your medieval times, when graphic imagery such as coats of arms
trust so they can steal your data. It is becoming a big issue in were used to distinguish between the nobility’s statuses.
the online world and needs to be addressed. This article will
discuss the initial project background of our new Online Fake
Logo Detection System.
Despite the conventional translation as ”word”, logos is not
I. I NTRODUCTION used for a word in the grammatical sense—for that, the term
Today,the world has witnessed massive computing power. lexis (, léxis) was used.[13] However, both logos and lexis
From Banking sector to Academic institutions, from defense to derive from the same verb légō (), meaning ”(I) count, tell,
corporate sector, people are extremely dependent on IT power say, speak”.[1][13][14]
which comes with massive energy consumption.[1][2][3] The Computer Vision Laboratory from New York University
has created a tool that lets you scan any logo and get the
Energy consumption of the Information and Communication findings of whether it is real or not. It’s basically some AI
Technology (ICT) sector has grown exponentially in recent software that allows you to compare the image of a random
years.We require a sizable database of archived photos in order file to your company logo.
to verify and distinguish between authentic and fake logos
in order to establish if a logo is legitimate or not. However,
this demands a lot of room in data centers, and maintaining B. Aristotle’s rhetorical logos
such data centers consumes a lot of energy..To determine
Following one of the other meanings of the word, Aristotle
if a logo is phony or real, each picture must be recorded
gave logos a different technical definition in the Rhetoric,
in a database, necessitating the use of several servers and
using it as meaning argument from reason, one of the three
data centers. Improving energy efficiency in data centers can
modes of persuasion. The other two modes are pathos (,
not only save costs, but also help mitigate the environmental
páthos), which refers to persuasion by means of emotional
impact of data centers.[4][5][6]
appeal, ”putting the hearer into a certain frame of mind”;[21]
and ethos (, êthos), persuasion through convincing listeners
This system was created to detect when a customer logo is
of one’s ”moral character”.[21] According to Aristotle, logos
being copied by someone else on social media or other web-
relates to ”the speech itself, in so far as it proves or seems to
sites without the customer’s permission so that the customer
prove”.[21][22] In the words of Paul Rahe:
can then take appropriate action.Our algorithm looks for nearly
identical logos in shape and design but differs in one or more
For Aristotle, logos is something more refined than the
details (a colour change, a few lines removed from an icon).
capacity to make private feelings public: it enables the human
Some logos, like the Starbucks logo, have many variations that
being to perform as no other animal can; it makes it possible
can go unnoticed
for him to perceive and make clear to others through reasoned
Identify applicable funding agency here. If none, delete this. discourse the difference between what is advantageous and
 The first kind depends on the personal character of the
speaker; the second on putting the audience into a certain
frame of mind; the third on the proof, or apparent proof,
provided by the words of the speech itself.
—Aristotle, Rhetoric, 350 BC[26]
C. Modern Day Logo’s
The modern era of logo design began in the 1870s with the
introduction of the first abstract logo, the Bass red triangle.
Logos became essential for brands to be memorable to poten-
tial customers as a result of the introduction of color printing
and the advertising industry. The author uses a case study to
show that fake websites can look like they are real companies,
so there is a need for websites that work like LogoMotive. The
article also provides statistics of how many more companies
want their domains monitored with LogoMotive than without
(Hout, 2022).[3]

The Internet is a vast domain that remains largely unex-

plored and poorly understood. As a result, scammers are able
to take advantage of the anonymity and ignorance that prevails
online. Scams are present in all shapes and sizes, but one of the
most pervasive forms of phishing activity is scam attempts on
e-commerce websites that use brand logos as bait. This form of
counterfeiting can be challenging to detect by consumers as it
Fig. 1. Aristotle, 384–322 BC.
tends to blend seamlessly into legitimate websites (Hesselman,
[35]
2022).[12]

Fake news is a new phenomenon that has been recently

what is harmful, between what is just and what is unjust, and
showing up all over the internet. It’s content that is created
between what is good and what is evil.[6]
and shared on social media sites like Facebook and Twitter in
order to deliberately mislead or misinform readers. This can
Logos, pathos, and ethos can all be appropriate at different be done for a variety of reasons with the most common being
times.[23] Arguments from reason (logical arguments) have moneymaking schemes, political influence, and online scams
some advantages, namely that data are (ostensibly) difficult (Zhang Ghorbani, 2020).[13]
to manipulate, so it is harder to argue against such an argu-
ment; and such arguments make the speaker look prepared Falsehoods are in the air today, with opinions and emotions
and knowledgeable to the audience, enhancing ethos.[citation often driven by misinformation. Some of these false reports,
needed] On the other hand, trust in the speaker—built through however, can be detected through a process called multimodal
ethos—enhances the appeal of arguments from reason.[24] multi-image fake news detection. Trying to detect fake news
from a single image is difficult enough. Humans rely on per-
Robert Wardy suggests that what Aristotle rejects in sup- ception and context cues that machines don’t fully understand
porting the use of logos ”is not emotional appeal per se, but when analyzing images for authenticity (Giachanou et al.,
rather emotional appeals that have no ’bearing on the issue’, 2020).[14] There is a significant disparity in the frequency at
in that the pathē [, páthē] they stimulate lack, or at any rate are
not shown to possess, any intrinsic connection with the point which and type of behavior that online reviewers are observed
at issue—as if an advocate were to try to whip an antisemitic to engage in with reviews that have been found to be fake,
audience into a fury because the accused is Jewish; or as verses reviews that have been validated as authentic. As a
if another in drumming up support for a politician were to result, there is an obvious need for personalization of review
exploit his listeners’s reverential feelings for the politician’s detection mechanisms. A system should be able to take into
ancestors”.[25] account the reviewer’s individual behaviors and correlate them
with the reviewer’s verbal and nonverbal cues (Zhang et al.,
2016).[15]
Aristotle comments on the three modes by stating:
Of the modes of persuasion furnished by the spoken word In the emerging internet-oriented economy, online recruit-
there are three kinds. ment fraud is one of the major issues for employers. This arti-
cle presents an intelligent model for online recruitment fraud ”The current approach is limited to detecting logos in images.
detection in a paper-less environment (Alghamdi Alharby,
2019). The proposed model has four modules: data collection,
data preprocessing, classification and prediction. The result of However, as research progresses, this approach is expected
the analysis indicates that the proposed method can effectively to scale up to more general image applications. Even though
detect the occurrence of fraud from its recruitment data to it’s called ’fake’ or ’logo’ detection system- This system can
identify potential suspects.[16] be used against any logo by anyone. This project was an
attempt to build a tool that allowed people to accurately detect
This article will discuss a number of post-production tech- logo counterfeits using only open-source software. Now that
niques that can be used to remove logos from video footage. we’ve created and demonstrated how the tool works, we intend
The first technique is to use Optical Flow Tracking, which to continue building on this work and see where it can lead.
aims to detect any logos or other marks in the scene, and This logo detection system will be useful for many companies.
if so, places these at fixed positions on the screen for the
remainder of the clip with a slight transform applied. The
second technique is to use Advanced Composition Modes in The proliferation of phishing attacks in recent years has
Premiere Pro, which looks for any rectangular objects within a made it imperative for organizations to have an effective way
specified colour range and removes them automatically (Yan, of detecting bogus websites and emails (Bozkir Aydos, 2020).
Wang Kankanhalli, 2005).[17] However, brand owners often find themselves in a dilemma
as they are reluctant to provide their logos publicly due to
This paper present a region-based convolutional neural net- security-related concerns. This paper will aim to address this
work architecture to detect logos in photos taken from different issue by explaining a new technique called HOG based logo
regions. The proposed method can detect almost any kind of detection scheme that offers brand owners the protection they
logo (text, geometric shapes e.g. , circle, triangle and line) in need while providing website operators with assistance in
images taken from regions of different sizes, with pixel-wise detecting fraudulent or malicious content.[2]
accuracy at a low computational cost and high speed (Bao et II. LITERATURE REVIEW
al., 2016).[18] The growing and massive production of visual data by
businesses and institutions, as well as the growing popularity
An automated script for logo detection and extendibility of social systems Graphics logos are a type of visual object
is described. The system detects if an image contains a logo, that is extremely important for determining the identity of
outputs the probability of a logo being present, and provides an something or someone. Logos are graphic creations that either
estimate of the size of the logo. Results are obtained through recall real-world objects, highlight a name, or simply display
the use of a convolutional neural network pre-trained on Im- abstract signs with strong perceptual appeal.
ageNet dataset with Tensorflow as the underlying framework
(Li et al., 2014).[19] The majority of trademark recognition research focuses
on the problem of content-based indexing and retrieval in
D. Evolution of a logo logo databases, with the goal of assisting in the trademark
registration process. The image acquisition and processing
The role of logos in our culture is evolving as a result
chain is controlled in this case so that the images are of
of technological advancements. We can see how logo design
acceptable quality and not distorted.
has evolved from complexity to simplicity, which is reflected
in the visual overload we’ve experienced as a result of our
A general system for detecting and recognizing logos in
increasingly complex lifestyles. The goal of creating a unique
photos captured in real-world contexts must meet a variety of
and simple logo mark that is both distinguishable and easily
constraints. On the one hand, conformance to a wide range
recognizable should be the goal of designing an authentic logo
of geometric and photometric modifications is necessary to
for a brand. A good logo today is adaptable in both design
meet all potential image/video recording conditions. Because
and application and can ideally stand alone. In today’s world,
logos are not captured in isolation in real-world photos,
the simpler the logo, the more easily recognizable it is.
logo detection and recognition should be resistant to partial
occlusions. Simultaneously, if we wish to detect malicious
tampering or retrieve logos with some local peculiarities, we
The computer vision algorithm identifies three key features:
must also ensure that minor deviations in local structures be
shape, contour, and luminance (which can be thought off
stored in the local descriptor and be enough differentiating for
as tonal value). These features must match in order for the
recognition.
machine to flag it as a fake or real. If these criteria are met,
then it will identify the file as genuine. The algorithm runs on A. Scammers’ use of trademarks to defraud consumers
the computer’s graphics processing unit (GPU) and resides in Logos give a website a familiar feel and promote trust.
the cloud, so it can be run on any device with a Web browser. Scammers take advantage of that by using well-known or-
ganizations’ logos on malicious websites. Unsuspecting In-
ternet users see these logos and think they are looking at
a government website or legitimate webshop, when it is a
phishing site, a counterfeit webshop, or a site set up to spread
misinformation. We present the largest logo detection study
on websites to date. We analyze 6.2M domain names from
the Netherlands ’ country-code top-level domain .nl,

In two case studies to detect logo misuse for two orga-

nizations: the Dutch national government and Thuiswinkel
Waarborg, an organization that issues certified webshop trust Fig. 4. Peugoat- Original vs Fake Logo
marks. We show how we can detect phishing, spear phishing, [51]
dormant phishing attacks, and brand misuse. To that end,
we developed LogoMotive, an application that crawls domain
names, generates screenshots, and detects logos using super-
vised machine learning. LogoMotive is operational in the .nl
registry, and it is generalizable to detect any other logo in any
DNS zone to help identify abuse.[1]
B. Fake Logo Examples
Fake logos disguised as advertisements are unsolicited
emails that claim to be advertisements (i.e., they pretend to be
from a particular company), but they contain hidden phishing
links that direct you to a website where the user’s personal
information could then be accessed without their knowledge Fig. 5. Volvos- Original vs Fake Logo
or permission. [51]

actual vendor instead of an individual looking for personal

information. This is done so that by the time the user opens
the email, there is less suspicion about its true nature, allowing
for more details about the target person to be gathered before
any action is taken. There are also scams that use a celebrity’s
likeness for the fake logo, making it look as if the celebrity
were endorsing a particular product. Some scammers even use
false logos from a company that is like the company they’re
trying to scam to fool users into thinking the scammer is from
that company.
Fig. 2. Bmw’s- Original vs Fake Logo
[51] C. Reasons
Scammers create fake logos for many reasons, but the most
common is personal gain (usually financial). To achieve this,
scammers need your personal information, such as credit card
numbers and bank account information. Personal information
can be sold or can be used to commit fraud on your behalf.
Scammers can obtain your personal information by getting
you to give it to them by clicking on a link in their email. This
is known as phishing when someone pretends to be someone
else to get you to do something you normally wouldn’t do.
There are several ways that scammers create fake logos
and many variations of scams that use them, but most of the
Fig. 3. Koeinseggg- Original vs Fake Logo scams use logos designed to look like they come from an
[51] official company to trick people into giving away personal
information. Often, they will even claim the logo is ”affiliated
Some scams send emails with fake logos disguised as scam with” the company or use a variation of the actual logo to
vendors to make it appear that the email was sent from an make their emails appear more legitimate.
D. Awareness The qualitative methodology of online logo detection system
To help you avoid being targeted by fake logos and scams, measures each logo on a scale from 1-5 and then calculates
it is important to learn to identify the different types of logos scores for each one. It can detect, compare and rank logos by
that could be used in phishing emails. comparing them with the highest scoring ones in their class.
To protect yourself from becoming a victim of fake logos
and scams, you must know what you’re looking for so that This is made possible because of the use of deep learning
you can spot it in time to avoid any loss. technique which has been used to learn and determine the
A wide range of techniques exists to detect and prevent following:
phishing attacks, also known as fake logo attacks.
Anti-virus software and spam filter programs are used to B. Proposed System
identify malware and prevent it from being delivered to the
We provide a unique technique for logo identification and
user’s computer. However, sometimes these programs do not
recognition based on the notion of ”Context-Dependent Sim-
detect the malicious content, or it may be delivered because
ilarities” in this research. The matching is done in this pro-
of an error in the program. In this case, anti-spam software is
cedure by separating the logo picture into rows and columns.
used to filter out known phishing emails before they get into
When this procedure is completed, the matching will be quite
the inbox. Anti-spam software uses both content filtering (to
exact. The approach has been demonstrated to be very success-
identify phishing emails by keywords) and reputation filtering
ful in meeting the criteria of logo detection and identification
(to identify phishing emails by email address).
in real-world photos. The likelihood of successful matching
and detection is high.In Figure 6 there is a clear point of view
If phishing emails are detected after they have been opened
on how does the design flow will be represented.
by users, effective anti-virus software and email filtering can
identify and delete the malicious code before it can do any • Logo Input: This is the input module of the system, where

damage. Once a computer has been infected with malicious logos are submitted for analysis. These logos may come
software, anti-virus software may be able to detect and remove from a variety of sources, such as websites, social media
it if it is updated regularly with malware definitions. However, platforms, or e-commerce sites.
anti-virus software may fail to detect phishing attacks because • Preprocessing: In this module, the logos are preprocessed

of how they are designed. and prepared for analysis. This may involve resizing the
logos to a standard size, removing background noise
III. S YSTEM OVERVIEW or clutter, or performing other preprocessing steps to
The system has two algorithms built into it. The first one improve the quality of the logo for analysis.
detects the presence of a slight colour change on logo ’A’ • Feature Extraction: In this module, the system extracts
while looking at two logos with the same shape but different relevant features from the logo that will be used for anal-
colours and detail differences in the details of a few lines in ysis. These features may include color, shape, typography,
an icon. The second algorithm is a simple BING image search and other characteristics of the logo.
to see if logo ’B’ shows up on websites other than that of the • Machine Learning Model: This is the core of the system,
customer that owns logo A. where a machine learning model is used to analyze the
features of the logo and compare them to a database of
A. Methodology known legitimate logos. The model is trained to recognize
All the logos in the world can be found everywhere. They’re patterns in the features of the logo that are indicative of
used to represent everything from clothing brands to coffee authenticity.
shops, from companies to politicians. And more than anything • Decision Making: Based on the output of the machine
else, people want their logo identity recognized on the digital learning model, the system makes a decision about the
web. But there’s a hidden side of these logo images that you authenticity of the logo. If the logo is determined to be
could never previously have seen: they might also be quite fake, it is flagged for further review or removed from the
difficult to spot online! As computer-aided designers move system.
towards increasingly complex and less-scannable designs that • Output: The final output of the system is a decision
resist easy reproduction by machine, it becomes important for about the authenticity of the logo, along with any relevant
researchers to find innovative ways logotypes can be identified information or recommendations for further action. This
online using qualitative methods or algorithms. And this is output may be presented to a user or incorporated into
where researchers are finding that things are changing. other systems or processes.
Overall, this system uses a combination of machine learning
and image processing techniques to analyze the features of C. Context-Dependent Similarity Algorithm
logos and make decisions about their authenticity. By contin- Let SX = x1,..., xn, and SY = y1,..., ym be the list of interest
ually updating and refining the machine learning model and points extracted from a reference logo and a test picture,
the database of legitimate logos, the system can improve its respectively (the value of n, m may vary with SX , SY ).
accuracy and effectiveness in detecting fake logos. We use the context and similarity design definitions to offer
 n→+
→ 0 and P KYJ |X
m
_
j _=J
KYj |X
n→+
→ 1.
Fonts that are most commonly found in online logos, A
method to detect logos without any text, A way to score every
logo as they come through; giving an objective standard that
allows for easy comparison. Deep learning is a technique that
uses a neural network to achieve high accuracy’s and is able
to learn through experiences.
Fig. 6. Design flow of fake logo detection
It is used in this system to recognize text, logos and creates
its own neural network. With this, the system can process
images at high accuracy rates, much higher than chance. It uses
a novel matching approach for logo detection. The algorithm the information provided by the user while they score images
utilized in this case is as follows: on their own scale of 1-5, which allows for the comparisons
between classes of logos to be made. The logo detection
Algorithm 1: CDS Logo Detection and Recognition. system, which has now become the logo recognition system,
is then able to detect a group of previously unseen logos with
for i ← 1 to n do
high accuracy. This is mainly due to the fact that there are
Compute the context of xi, given _,Na,Nr;
clear and objective standards that the logos must follow and
for j ←1 to m do
it makes comparisons between logos much easier. The use of
Compute the context of y j, given _,Na,Nr;
deep learning also allows for this system to process images at a
Set t ←1, maxt ←30;
minimum of 90 percent accuracy because deep learning is used
repeat
in places where there is no text to help guide it, for instance
for i ←1 to n do
text such as ”New York” repeated throughout the image does
for j ←1 to m do
not need to be relearned because it has already been learned
Compute CDS matrix entry K(t )
in previous images.
xi ,y j, given , ;
Set t ←t + 1; IV. P ROJECT U NIQUENESS
until convergence (i.e., _ _ A. Work Flow
K(t ) K(t1)_ This is a project that works on the detection of fake online
_2 _ 0) OR logos. This system allows users to upload the image they
t > maxt ; want to verify whether it is legit. This system process the
K ← K(t ); uploaded images by detecting various type of areas in it with
for i ← 1 to n do unique algorithms and comparing its result with a database of
for j ←1 to m do information that was previously generated by this same system
Compute Ky j |xi and reporting back to the user in multiple regions as well as
← pointing out what kind of errors are found on the images such
Kxi ,y j as missing/wrong words/ edges, etc.
_ms
=1 Kxi ,ys B. Data Gathering
; Because of Instagram’s picture-centered nature and pop-
A match between xi and y j is declared iff ularity among marketers, I elected to concentrate my data
Ky j |xi _ _ms gathering (i.e., image collecting) efforts there. Unfortunately,
_=j Kys |xi ; the Instagram API just undergone significant modifications
if number of matches in SY > |SX | then that severely restrict access. As a result, I elected to man-
return true i.e. logo detection ually extract photographs from various Patagonia-sponsored
else athletes’ and Patagonia retailers’ Instagram profiles. I utilized
return false; JavaScript on a Google Chrome console to accomplish the
Provided that 1/q manual scraping. It wasn’t perfect, but it allowed me to rapidly
collect photographs and begin training my model!
 Because this was a supervised problem, I looked through where it will be used to detect fake logos in real-world
all of the photographs I downloaded and organized them into scenarios.
two folders: 1) possesses a Patagonia logo (logo), or 2) lacks • Continuous monitoring and improvement : The system
a Patagonia logo (no-logo). would need to continuously monitor and improve by
updating dataset and retraining the model to adapt to new
C. Model Development types of fake logos and improve its accuracy over time.
I spent the most of my time iterating on the logo detection
model. The key tasks were to alter the out-of-the-box Inception V. R ESULTS AND D ISCUSSIONS
v3 model and related retraining script in TensorFlow, retrain In the process of developing a logo detection system, I
the new model, analyze model performance, and adjust model faced two major problems: A lack of expertise on how to
parameters as needed. train a model in some required algorithms, and the difficulty
I made the following significant changes to the out-of- in generating images with non-standard resolutions. These
the-box Inception v3 model and related retraining script in problems forced me to develop my own implementation for
TensorFlow: recognizing logos with respect to their difference in resolu-
Retrained the final model layer to distinguish between logo tions. This implementation is based off using convolutional
and no-logo. neural networks. By implementing this algorithm, I achieved
an accuracy rate of over 60 on the MNIST dataset and the
test data provided by Yahoo logo detection. This system had
a training time of only 37 minutes which is much faster than
other solutions that use 40 hours or more for training . I think
this system can be used to automatically detect logos in a
website. Fig. 8, 9, and 10 show how a single fake logo’s
analysis produces a portrayal of the type of fraud and how
much it covers in a pie chart. They also show the varying
outcomes when data base management system is used to
optimize and tame the findings.

Fig. 7. Architectural flow for Fake logo detection

The logo class of photos was up sampled to increase

accuracy and recall for the imbalanced class.In fig 7 the
architectural diagram can we viewed as The design flow for
a fake logo detection system would involve several steps,
including:
• Data Collection: The first step would be to gather a
dataset of both genuine and fake logos. This dataset
would be used to train and test the fake logo detection
system. Fig. 8. illustrating how a bar graph and a pie chart are represented when
analyzing a false logo.
• Feature Extraction: Once the dataset is collected, the next
step would be to extract features from the images that can
be used to differentiate genuine logos from fake ones.
This could include features such as shape, color, texture,
and typography.
• Model Training: Using the dataset and extracted features,
the next step would be to train a machine learning model
to recognize the patterns and characteristics of genuine
logos and fake logos. This could be done using techniques
such as deep learning or computer vision.
• Model Evaluation: After the model is trained, it would
be evaluated on a separate dataset to test its accuracy in
detecting fake logos. This could be done by comparing
the model’s predictions to the true labels of the logos in
the evaluation dataset.
• Model Deployment: Once the model is deemed accurate Fig. 9. Degree of fraud detection with optimization.
enough, it can be deployed to the production environment
 Prediction Confidence Score Mean Confidence Average Confidence
Real 0.97 0.94 0.93
Fake 0.89 0.86 0.88
Real 0.93 0.91 0.92
Fake 0.87 0.85 0.86
Real 0.99 0.97 0.98
TABLE II
M EDIAN AND AVERGAE SCORES

of the five unique logos about their legitimacy. The confidence

score, which goes from 0.0 to 1.0 with zero being the lowest
Fig. 10. Degree of fraud detected with optimization and time taken by DBMS
and 1.0 being the greatest, allows us to evaluate an image’s
credibility. Table 2 shows each logo is assigned a prediction of
either ”Real” or ”Fake” based on the output of the fake logo
detection system. The confidence score is a value between 0
and 1 that indicates the system’s confidence in its prediction.
A high confidence score (e.g. 0.99) indicates that the system is
very confident in its prediction, while a low confidence score
(e.g. 0.50) indicates that the system is less certain.

This above table shows average scores of the logos classified

into categorised of size and colors
With the training dataset, I achieved a test accuracy of
66 by using Tensorflow and the implementation of three
different neural networks: AlexNet, GoogLeNet, and VGGNet.
These are convolutional neural networks that are used to
classify images into categories based off different features.
A convolutional network has neurons arranged in multiple
layers. It uses a sliding window that slides through all of the
pixels in an image when trying to identify the object in it.
Fig. 11. Tensor Flow of fake logo detection Each layer in the network calculates some representation of
the input from its previous layer, then applies a non-linear
transformation before passing it on to subsequent layers.
TensorFlow, a potent open-source machine learning soft-
ware framework created by Google, is discussed in Fig.11. It
may be used to create and train models for many different
tasks, including as object identification, picture classification,
and natural language processing.
TensorFlow is probably used in a ”fake logo identification
system” to train a model that can distinguish between actual
and fraudulent logos. This may entail training the algorithm
on a dataset of authentic and fraudulent logos before using
it to categorize previously undiscovered logos. Convolutional
neural networks (CNNs), which are frequently employed for
image classification applications, would probably be used in
the procedure. Table 1 displays the confidence scores for each

Logo Result Confidence Score

Fig. 12. scattered plot
1 Real 0.95
2 Fake 0.70 Fig.12 displays the outcomes of an online system for
3 Real 0.99 detecting false logos on a dataset of logos. The dataset’s logos
4 Fake 0.60 are shown on the x-axis, while each logo’s confidence scores
are shown on the y-axis. The confidence scores are displayed
5 Real 0.80 as a line graph with the confidence scores for the ”Real” logos
TABLE I
TABLE SHOWING CONFIDENCE SCORES FOR 5L OGOS shown on the line and the confidence scores for the ”Fake”
logos shown on the bars.
 In this graph, the confidence scores for the ”Real” logos are pictures by training it on additional data. Localization of logo
generally higher than those for the ”Fake” logos, indicating detection (i.e., where, if any, is a logo in this image?) Combine
that the fake logo detection system is more confident in its existing textual analytics with logo detection. Model selection
predictions for the ”Real” logos. The overall shape of the and hyper-parameter adjustment should be automated. Investi-
graph may also give insights into the performance of the gate various model designs. Contrast model performance with
fake logo detection system. For example, if the confidence those of existing logo-detection systems.
scores are consistently high or low across all of the logos,
this may suggest that the system is performing well or poorly, R EFERENCES
respectively. The results from a collection of logos utilizing
[1] Bhattacharya, Tathagata, et al. ”Capping carbon emission from green
data centers.” International Journal of Energy and Environmental Engi-
neering (2022): 1-15.
[2] Peng, Xiaopu, et al. ”Exploiting Renewable Energy and UPS Systems
to Reduce Power Consumption in Data Centers.” Big Data Research 27
(2022): 100306.
[3] Bhattacharya, Tathagata, and Xiao Qin. ”Modeling Energy Efficiency
of Future Green Data centers.” 2020 11th International Green and
Sustainable Computing Workshops (IGSC). IEEE, 2020.
[4] Bhattacharya, Tathagata, et al. ”Accelerating the Energy Efficient Design
of Traditional Data Centers Through Modeling.” 2022 IEEE Interna-
tional Conference on Networking, Architecture and Storage (NAS).
IEEE, 2022.
[5] Bhattacharya, Tathagata, et al. ”Performance modeling for I/O-intensive
applications on virtual machines.” Concurrency and Computation: Prac-
tice and Experience 34.10 (2022): e6823.
[6] Cao, Ting, et al. ”DDoS Detection Systems for Cloud Data Storage.”
2021 Third IEEE International Conference on Trust, Privacy and Secu-
rity in Intelligent Systems and Applications (TPS-ISA). IEEE, 2021.
Fig. 13. scatter plot visualize the results [7] Henry George Liddell and Robert Scott, An Intermediate Greek–English
Lexicon: logos, 1889.
an online fake logo detection approach are shown in Fig. 13’s [8] Entry at LSJ online.
graph. The logos in the dataset are displayed on the x-axis, and [9] . Owl.purdue.edu // Purdue Writing Lab ”Aristotle’s Rhetorical Situation
// Purdue Writing Lab”. Owl.purdue.edu. Retrieved 2022-03-16.
the confidence ratings for each logo are displayed on the y- [10] Cambridge Dictionary of Philosophy (2nd ed): Heraclitus, (1999).
axis. With the confidence ratings for the ”Real” logos shown [11] Paul Anthony Rahe, Republics Ancient and Modern: The Ancien
on the line and the confidence scores for the ”Fake” logos Régime in Classical Greece, University of North Carolina Press (1994),
ISBN 080784473X, p. 21.
shown on the bars, the confidence scores are presented as a
[12] Rapp, Christof, ”Aristotle’s Rhetoric”, The Stanford Encyclopedia of
line graph. It’s important to note that the results of a fake Philosophy (Spring 2010 Edition), Edward N. Zalta (ed.)
[13] David L. Jeffrey (1992). A Dictionary of Biblical Tradition in English
logo detection system will depend on the quality and variety Literature. Grand Rapids, Michigan: Wm. B. Eerdmans Publishing Co.
p. 459. ISBN 978-0802836342.
of the training data used to develop the system, as well as [14] Cambridge Dictionary of Philosophy (2nd ed): Philo Judaeus, (1999).
the performance of the system itself. As such, the results of a [15] Adam Kamesar (2004). ”The Logos Endiathetos and the Logos
fake logo detection system may vary and should be interpreted Prophorikos in Allegorical Interpretation: Philo and the D-Scholia to
the Iliad” (PDF). Greek, Roman, and Byzantine Studies (GRBS). 44:
with caution. 163–181. Archived from the original (PDF) on 2015-05-07.
[16] May, Herbert G. and Bruce M. Metzger. The New Oxford Annotated
CONCLUSION AND F UTURE D IRECTIONS Bible with the Apocrypha. 1977.
I was quite pleased with the results of my initial model and [17] Afroz, S., Greenstadt, R.: PhishZoo: detecting phishing
websites by looking at them. In: 2011 IEEE Fifth International
the web app’s basic functioning. But it was only the top of the Conference on Semantic Computing. IEEE, September 2011.
iceberg! Here are some potential future stages for this project https://doi.org/10.1109/icsc.2011.52
that I’ve been considering: [18] Bozkir, A. S., Aydos, M. (2020). LogoSENSE: A companion HOG
based logo detection scheme for phishing web page and E-mail brand
Understanding how such a system may be successful while recognition. Computers Security, 95, 101855.
yet safeguarding people’s privacy. To establish product/market [19] Hout, T. V. D., Wabeke, T., Moura, G., Hesselman, C. (2022, March).
fit, pitch use-cases to brands/marketers. Expand the product LogoMotive: detecting logos on websites to identify online scams-a TLD
case study. In International Conference on Passive and Active Network
prototype and get feedback Allow bulk image processing Measurement (pp. 3-29). Springer, Cham.
over a RESTful API. Integrate with Instagram’s (and other [20] Hesselman, C. (2022). LogoMotive: Detecting Logos on Websites to
social media picture services’) API to gather photographs Identify Online Scams-A TLD Case Study. In Passive and Active
Measurement: 23rd International Conference, PAM 2022, Virtual Event,
automatically. March 28-30, 2022: Proceedings (Vol. 13210, p.
[21] Zhang, X., Ghorbani, A. A. (2020). An overview of online fake
Enhance and improve the model/process Improve recall on news: Characterization, detection, and discussion. Information Process-
the ”has logo” class by include class weights in the loss ing Management, 57(2), 102025.
[22] Giachanou, A., Zhang, G., Rosso, P. (2020, October). Multimodal multi-
function, as proposed in this StackOverflow article. Increase image fake news detection. In 2020 IEEE 7th International Conference
the capacity of the model to generalize to a larger group of on Data Science and Advanced Analytics (DSAA) (pp. 647-654). IEEE.
[23] Zhang, D., Zhou, L., Kehoe, J. L., Kilic, I. Y. (2016). What online [51] In the translation by W. Rhys Roberts, this reads ”the proof, or apparent
reviewer behaviors really matter? Effects of verbal and nonverbal proof, provided by the words of the speech itself”.
behaviors on detection of fake online reviews. Journal of Management [52] Eugene Garver, Aristotle’s Rhetoric: An art of character, University of
Information Systems, 33(2), 456-481. Chicago Press (1994), ISBN 0226284247, p. 114. Garver, p. 192.
[24] Alghamdi, B., Alharby, F. (2019). An intelligent model for online [53] Robert Wardy, ”Mighty Is the Truth and It Shall Prevail?”, in Essays on
recruitment fraud detection. Journal of Information Security, 10(03), Aristotle’s Rhetoric, Amélie Rorty (ed), University of California Press
155. (1996), ISBN 0520202287, p. 64.
[25] Yan, W. Q., Wang, J., Kankanhalli, M. S. (2005). Automatic video logo [54] Translated by W. Rhys Roberts, http://classics.mit.edu/Aristotle/rhetoric.mb.txt
detection and removal. Multimedia Systems, 10(5), 379-391. (Part 2, paragraph 3)
[26] Bao, Y., Li, H., Fan, X., Liu, R., Jia, Q. (2016, August). Region-based [55] https://logos-world.net/nike-logo/
CNN for logo detection. In Proceedings of the International Conference [56] https://www.dailymail.co.uk/femail/article-7747625/Tricky-pictures-
on Internet Multimedia Computing and Service (pp. 319-322). challenge-players-spot-real-logo-fake.html
[27] Li, K. W., Chen, S. Y., Su, S., Duh, D. J., Zhang, H., Li, S. (2014). [57] https://raw.githubusercontent.com/ilmonteux/logohunter/master/pipeline.gif
Logo detection with extendibility and discrimination. Multimedia tools
and applications, 72(2), 1285-1310.
[28] Eggert, C., Winschel, A., Lienhart, R.: On the benefit of syn-
thetic data for company logo detection. In: Proceedings of the 23rd
ACM International Conference on Multimedia. ACM, October 2015.
https://doi.org/10.1145/2733373.2806407
[29] FBI: FBI Warns Public to Beware of Government Impersonation
Scams, April 2021. https://www.fbi.gov/contact-us/field-
offices/boston/news/press-releases/fbi-warns-public-to-beware-of-
government-impersonation-scams
[30] Fielding, R., Reschke, J.: Hypertext Transfer Protocol (HTTP/1.1):
Semantics and Content. RFC 7231, IETF, June 2014.
http://tools.ietf.org/rfc/rfc7231.txt
[31] FTC: How To Avoid a Government Impersonator Scam, April
2021. https://www.consumer.ftc.gov/articles/how-avoid-government-
impersonator-scam
[32] Goel, R.K.: Masquerading the government: drivers of government im-
personation fraud. Public Finan. Rev. 49(4), 548–572 (2021)
[33] Google: Google Public DNS (2021).
https://developers.google.com/speed/public-dns/
[34] Google Inc.: Certificate transparency. https://certificate.transparency.dev/
[35] Han, Y., Shen, Y.: Accurate spear phishing campaign attribu-
tion and early detection. In: Proceedings of the 31st Annual
ACM Symposium on Applied Computing. ACM, April 2016.
https://doi.org/10.1145/2851613.2851801
[36] Hesselman, C., Moura, G.C., Schmidt, R.D.O., Toet, C.: Increas-
ing DNS security and stability through a control plane for top-
level domain operators. IEEE Commun. Mag. 55(1), 197–203 (2017).
https://doi.org/10.1109/mcom.2017.1600521cm
[37] Hill, K.: The Secretive Company That Might
End Privacy as We Know It, January 2020.
https://www.nytimes.com/2020/01/18/technology/clearview-privacy-
facial-recognition.html
[38] Hoffman, P., Sullivan, A., Fujiwara, K.: DNS Terminology. RFC 8499,
IETF, November 2018. http://tools.ietf.org/rfc/rfc8499.txt
[39] Introna, L.D.: Disclosive ethics and information technology: disclosing
facial recognition systems. Ethics Inf. Technol. 7(2), 75–86 (2005).
https://doi.org/10.1007/s10676-005-4583-2
[40] Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization (2017)
[41] David L. Jeffrey (1992). A Dictionary of Biblical Tradition in English
Literature. Grand Rapids, Michigan: Wm. B. Eerdmans Publishing Co.
p. 460. ISBN 978-0802836342.
[42] Henry George Liddell and Robert Scott, An Intermediate Greek–English
Lexicon: lexis, 1889.
[43] Henry George Liddell and Robert Scott, An Intermediate Greek–English
Lexicon: legō, 1889.
[44] F. E. Peters, Greek Philosophical Terms, New York University Press,
1967.
[45] W. K. C. Guthrie, A History of Greek Philosophy, vol. 1, Cambridge
University Press, 1962, pp. 419ff.
[46] The Shorter Routledge Encyclopedia of Philosophy
[47] Translations from Richard D. McKirahan, Philosophy before Socrates,
Hackett, (1994).
[48] Handboek geschiedenis van de wijsbegeerte 1, Article by Jaap Mansveld
Keimpe Algra, p. 41
[49] W. K. C. Guthrie, The Greek Philosophers: From Thales to Aristotle,
Methuen, 1967, p. 45.
[50] Aristotle, Rhetoric, in Patricia P. Matsen, Philip B. Rollinson, and
Marion Sousa, Readings from Classical Rhetoric, SIU Press 1990), ISBN
0809315920, p. 120.