Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Preserving and Randomizing Data Responses in Web

Application using Differential Privacy
Rubia Fathima
REVA Academy for Corporate
Excellence REVA University
Bengaluru, India

Abstract:- The Data reflects Humans as most of the data is Individual user's Data that involves any personal private
online. Privacy for individual's data has become a prime information requires differential privacy to be applied. The
concern. Securing and preserving that privacy has been a most popular definition of privacy is differential privacy,
focus for long decades. In data analytics, machine learning coined as the new mathematical term came in 2004 in data
and data science analysis over users' private data utilize to privacy. It ensures publicly visible data does not make any
understand the individual user's responses to present data changes for a single individual if there are changes in the
publicly. The issue with private data presented online that dataset. It resolves issue by adding random noise to the
consisted of personal private data was sensitive and mechanism at work. The need for an increase in adding
confidential was a significant issue, so a particular group robustness in the form of noise, maintain meaning full data
of mathematicians and cryptographers came together to pattern, the mathematically stringent definition of privacy, and
resolve this issue by introducing the concept of Differential computation that would be rich in a class of algorithms that
Privacy. satisfy the definition of differential privacy. [1]

Differential privacy is assurance over information privacy Preserving Privacy and Security guarantees the data
without damaging the chance of having privacy risk by pattern has been spinning around the compliance process and to
including some amount of Random noise in the form of perform powerful collection and curation of data to apply
robust data to the original dataset. Differential privacy is appropriate policies for user's private data in any form
also a tool with an algorithm that helps maintain Privacy
by Preserving and Randomizing data responses— Differential Privacy Guarantees:
measuring the accuracy of statistical data by performing 1. The raw data holding individual responses will not be
analysis. To Perform this process of differential privacy, unauthorized access and does not need to be modified.
IBM developed an open-sourced algorithm called 2. Maintaining an individual's privacy will be valued over
Diffprivlib[1]. With this library, the project has created a mining important details from data.
Front-End Web application that can perform data analysis 3. Manage resilience to post-processing; output generated
that involves different mechanisms, models, and Tools. from the secret differential algorithm will not affect the
differential privacy of the algorithm. In other words, the
This project is an attempt to integrate all mechanisms, data analyst that does not have additional knowledge about
models, and tools involved in DiffPrivLib[1]. The primary the dataset cannot increase privacy loss by looking and
purpose of this paper is to showcase the work on thinking at the output of the Differentially Private
differential privacy that consists in developing a user- algorithm.[2]
friendly web application that can be open-sourced. This
application is designed in a python programming package The infield of research study on Machine Learning the
and will experiment with the dataset to perform the Scope has been growing, and flooding like a tidal surge, the in-
analysis to show the impact of differential privacy depth analysis in data privacy standard has emerged to
algorithms on different values on epsilon with accuracy differential privacy algorithms in the subjects of Cryptography
and privacy. and Security. IBM has presented great work by creating the
Differential Privacy library, a general purposed and open-
Keywords:- Differential Privacy, Python Programming, Open- sourced library for investigating, experimenting, and
Source Library, Data Science, Machine Learning, Data developing differential privacy applications in a python
Analytics. programming language called "Diffprivlib." The library
includes all host mechanisms, the building blocks of
I. INTRODUCTION differential privacy[1], and several other applications to
experiment in Machine learning and other data analytics tasks.
As exposure of electronic data over the internet has This project demonstrates the idea of differential privacy in a
become specific, detailed, and abundant, maintaining individual web application that can be useful for any data analyst or
privacy has become the top priority. So, securing these personal accountant to perform analysis in the form of mathematical
private data in the datasets has been processed in mathematical computations for the supervised dataset. This application solves
computation involved in differential privacy algorithms.

IJISRT21AUG402 www.ijisrt.com 139

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
the problem of having cost-effective results related to privacy 2. Data privacy violations such as breaches also damage the
as an open-source web application. reputation of business "2020 Cost of a Data Breach Report"
by IBM[4] states that lost business due to diminished
A. Scope of Work reputation was the most significant cost factor of a data
The main scope of the work is to develop a front-end web breach with a yearly average of $1.52 million. The report
application developed using digital technologies such as also states that customer's personally identifiable
AI/ML, Cryptography[3], and Security. The study and information was the costliest data type to compromise in the
experiment related to this project can help the enterprise in the data breach they studied.
following manner 3. Differential privacy enables trade to share their information
1. The Web application developed for IBM's Differential securely with other organizations to collaborate with them
Privacy library having covered mechanism, model, and without gambling their customer's privacy
tools. 4. It is resistant to privacy attacks based on supplementary
2. Application developed can perform analysis that is free and information to prevent possible linking attacks on de-
cost-effective identified data.
3. "Diffprivlib" provides an extensive collection of B. Abbreviations and Acronyms
mechanisms, the fundamental building blocks of differential
privacy that handle the addition of noise. The Parameter Abbreviation Long Form
used to set this value is denoted as epsilon (ε) to the dataset; GDPR General Data Protection Regulation
ε controls so much noise or randomness to a raw dataset. CCPA California Consumer Privacy Act
4. As the application is available opensource can be utilized to HIPAA Health Insurance Portability & Accountability Act
perform experiments by small-size companies to high-level GLBA Gramm-Leach-Bliley Act
companies PPA precise poverty alleviation
5. Helps non-technical person or provides ease of work for a PATE Private Aggregation of Teacher Ensembles
data analyst who can perform different computations and NIST National Institute of Standards and Technology
understand the data DP Differential Privacy
6. Experiments were performed only on the supervised
AI Artificial Intelligence
dataset.
ML Machine Learning
7. For Accountant, Data analysts can perform computations to
SQL Structured Query Language
understand if there is any Privacy Leakage.
PII Personally, Identifiable Information
Differential privacy is becoming a significant factor to
business because: II. LITERATURE SURVEY
1. It helps businesses analyze leakage and maintain data
privacy that can comply with GDPR and CCPA without The project aims to have deep analysis in research for
determining the ability to understand their customer differential Privacy has become a de facto standard to maintain
behavior. While not complying with regulations can result and guarantee privacy as it has become one of the hot research
in a severe and heavy number of fines. As per the 2021 topics in the field of healthcare especially. From privacy
recent report[4] from international law firm DLA Piper's engineering, there have been many new programs held to
cybersecurity and data protection team has been infringing understand the performance of the differential privacy tool in
with €273 million of fines for not complying with GDPR Machine Learning[5] and Data Science[6]. The Review of
since May 2018. These fines are a drop in the bucket if they background theory and existing literature, tools from
start considering the level of GDPR compliance and the renowned journals/technical reports/websites showcases the
European economy's size. The below-given figure shows limitations and advantages of those tools and previous work by
the total value of GDPR fines imposed from 25th May 2018 few researchers that must support the problem formulation on
to 2020 appears to have increased as per the country's check the deployment of differential privacy. NIST has a list of de-
for GDPR compliance with more comprehensive and identification tools in the ranking challenge from created and
automated approaches.[4] The below-given Figure1 shows during the Privacy Engineering program[7]; each application
the Total values of GDPR fines. has research papers that explained limitations in the market.
There are around 23 de-identification tools in the list of NIST.
Most open-sourced tools are applied with limited edition and
paid versions come with fully functional modules but are not
cost-effective.

Hence, the problem stated above from the literature

survey explains the limitations and drawbacks of few tools;
therefore, the problem statement that this project is trying to
explain is addressing the same and enable in the future with an
open-source web application from IBM called Diffprivlib.
This tool is by a non-IT person that performs administrative
work or general staff to generate the output for getting privacy
Fig 1: Total value of GDPR Fines Imposed since May 2018[4]
in data, perform automatic computations on non-Zero count,

IJISRT21AUG402 www.ijisrt.com 140

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Mean, Variance and Standard Deviation, Histogram and train From the research and literature review on the available
model to get values that any generalist has understood in the tool for Differential Privacy, the inference is that most of the
form of a graph. DP application is just data masking and anonymization tools
but to preserve and maintain data accuracy without making
Based on the above literature study, it is clear that this considerable changes epsilon. Randomizing responses[8]
project can address whitespace with a survey on different De- guarantees the privacy of the data and protects it from different
identification tools performed by NIST[7]. The available tools attacks like Reconstruction attacks, Linkage attacks, etc. The
have fully functional differential Privacy computations with main aim of data curator/creator is to develop front-end
paid versions; otherwise, if the application is open source, the applications and perform the proper computation for precise
service is limited to have used only a limited mechanism datasets.
involved. This project efficiently implements all available
tools, models, and mechanisms in Diffprivlib with a IV. PROJECT METHODOLOGY
lightweight python framework on the web application. The
non-IT person with minimum data science knowledge can The project was developed based on SDLC phases and
utilize the tool to perform differential privacy computations followed by Data Science or Machine Learning lifecycle, and
and mechanisms. While looking at other tools in the market, the methodology contains six different points that followed a
there may be a precedence but maintaining resilience and Waterfall model approach. Below the given diagram, Figure 2
sustainability. has shown project methodology.

III. THE OBJECTIVE OF THE STUDY

The purpose of the capstone project is to have a well-

integrated user-friendly web application that is high in
resiliency and sustainability using IBM's Diffprivlib library,
that can perform computations based on digital technologies
such as AI/ML, Cryptography, and Security. The objectives of
this work are below
1. The primary objective of the differentially private algorithm
is to measure privacy and reduce risk.
2. The secondary objective is to investigate different datasets
and experiment with developed web applications in
differential privacy.
3. Understand the importance of mathematical noise to the
data can help preserve an individual's privacy and
confidentiality.
4. Enables organizations to customize the privacy policy to
prevent attacks like Linkage and reconstruct the original
Fig 2: Project Methodology
data.
A. Process Steps
5. The user's learning dataset will have to improve with
1.
Observation and understanding of the parameters of IBM's
2. Gathering Data: During the first phase, the dataset is
open-sourced differential privacy library along with
captured by uploading the dataset
machine learning and other data analytics task.
6. With differential privacy, algorithm parameter Epsilon (ε)
3. Data Preprocessing: The data is analyzed to understand
changes the accuracy and increases the dependency of
the use case; in the case of Differential privacy, the most
Epsilon ε over privacy.
looked into is identifying the critical column with personal
private data of the user or data owner. They are here
The novelty of this project is to provide the computational
preprocessing the data for feature extraction, feature
algorithm that performs sampling of the dataset so that original
analysis, and feature visualization.
data can be preserved with post-processing and retained in final
results. The tool that the originality of preserves data.
4. Train Model: For the Diffprivlib[2] Library models that
classification into supervised and unsupervised learning;
From the above Scope and objectives, an open-source
the below-given table has provided a sort between
library that develops will support many mechanisms, models,
different types of models, for open-sourced Web
and tools related to Machine learning. The main point to be
application has been implemented with integrated
considered is applying the algorithm so that data privacy
differential privacy and currently with limited to
maintains the same original dataset without changing patterns.
supervised learning and only classification model
Making data robust by adding some noise can be controlled by
GaussianNB, Linear regression, and Logistic Regression is
Epsilon and data measures with accuracy. Calculates non-Zero
applied. It is the responsibility of the user to choose the
count to understand the truthfulness of the data for
appropriate model, keeping the uploaded dataset in mind.
mathematical computations like Mean, Standard Deviation,
Variance of the dataset.

IJISRT21AUG402 www.ijisrt.com 141

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
5. Test Model: The selected model has to follow a specific A. Data Components:Diffprivlib is of four major components
mechanism included in Diffprivlib[2] Library with Web 1. Mechanisms: These are the building block of differential
application has captured the Analysis for supervised privacy uses in all models that actualize differential
dataset having Classification models such as GaussianNB, privacy. Mechanisms have little or no default settings for
Linear regression, and Logistic Regression. With each use use by experts implementing their models. They can be that
case, test a model by choosing a model among as it may be utilized exterior models for partitioned
GaussianNB, Linear regression, and Logistic Regression examinations, etc.[11]
and then choose the column to that column has to be target 2. Models: This module includes machine learning models
and display. with differential privacy. Diffprivlib currently has models
for clustering, classification, regression, dimensionality
6. Model Deployment: The application is deployed on a reduction, and preprocessing.
localhost python-based server using version 3.Code is also 3. Tools: Diffprivlib comes with several nonexclusive devices
available on private GitHub account It is of a lightweight for differentially private information analysis. Data
Python-based Framework called Flask. Diffprivlib[2] by incorporates differentially private histograms, taking after
default comes with functionality and familiarity of NumPy the same arrange as NumPy's histogram function.[12]
[17] and Scikit-learn [14] packages, meaning functions and 4. Accountant: The BudgetAccountant class can track
models are instantly recognizable, with default parameters security budget and calculate add up to security misfortune
ensuring accessibility for all. The web application is with utilizing progressed composition techniques.
python libraries such as NumPy, Scikit-learn, and SciPy
packages. The function flow diagram in Figure 4 explains the
process utilized for developing the robustness in data using a
7. Model Monitoring: The Web application over differential Differentially privacy algorithm in the below-given chart.
privacy algorithm involves data visualization that performs
and outputs a Mean, Variance, and standard deviation over
the dataset. At the same time, it can also perform a non-zero
count between Non-Private and Private data to measure the
truthfulness of the algorithm.
Account Setup Data Summarizer Visualization Train ML Model
Overall, the concept of the project is to manage and • Select dataset • Non Zero • Histogram Plots • Choose model -
• upload Dataset • Mean GaussianNB, Linear
perform the different privacy mechanisms involved in • Var and Logistic Regression
• Standard Deviation • Choose Target Column
Diffprivlib Library[9]. Still, while this project performs only
for the supervised dataset for classification and regression
model, in future analysis, this project will improve the
deploying mechanism for unsupervised raw data to perform
using the K-means algorithm.[10]

V. RESOURCE REQUIREMENT SPECIFICATION Fig 4: Data Flow diagram for DP web Application
The below-given Figure 3 shows diagram explains the
B. Process Overview Steps
process utilized for developing the Robust Differential privacy 1.Account Setup: Select the account to set up a temp session
data post-processing
for a data analyst or accountant. Setup budget account
1.Select the data set that would perform sampling
2.preview from the data selected
3.select the personal private data that is to mask
4.Upload-Epsilon value by Choosing epsilon Value
2.Data summarizer: Display the data of Private and Non-
Private and perform comparative analysis method and display
in the form of a graph having checked with accuracy
1. Non-Zero Count: This shows the number of non-zero
elements for each column in the dataset. The screenshot in
the next chapter clearly shows the change in the data
distribution of select columns of the dataset involved in the
differential private methods.
2. Mean: This shows the relative change in mean values of
different columns between the non-private/standard method
and differentially private counterpart.
3. Variance: This shows the relative change in variance
Fig 3: Process flow and modules in Differential Privacy values of different columns between the non-
private/standard method and differentially private
counterpart.

IJISRT21AUG402 www.ijisrt.com 142

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
4. Standard Deviation: This shows the relative change in The Software architecture of Python programming,
traditional deviation values of different columns between above given architecture, performs the analysis divided into
the non-private/usual method and differentially private parts for ease of understanding.
counterpart.
3. Visualization - The histogram plots of both non- On Back End
private/normal plotting functions and the Differential 1. Aggregate Data and Process: Dataset used that can used
Private plotting methods. Mechanism, Model and tools to perform analysis
4. Train ML Model - Apply machine learning principles on 2. Analyze and Sanitize data: post-Processing involves the
the dataset to train the model and obtain a comparative analysis to perform a test on the training data for the
report of the performance of non-private/standard and the different values of epsilon
unique private ML models. Training a model to analyze the
test dataset On the Front End
a) Choose Model: This option is to choose one of the many 1. Perform computations like non-Zero count, Mean, var and
models implemented with integrated differential privacy. SD
2. Test accuracy by comparing with Private and Non-Private
Currently, the below models: data
1.Gaussian Naive Bayes Model 3. Display the data for different ranges of Epsilon Values
2.Logistic Regression
3.Linear Regression A. Low-Level diagram:
Flowchart to explain each module involved in the web
The above given includes both classification models a application are Accountant, Mechanism, Model, and Tools
well as regression models. It is the responsibility of the user to
choose the appropriate model, keeping the uploaded dataset in 1. Pre-Processing Data given flowchart provides data flow
mind. for accountant module to show, so the preprocessing of the
dataset in Figure 6 involves applying mechanism like
b) Choose the Target column. This option uses to set the GaussianNB to perform aggregation of the dataset. The
dependent or the Y-column involved in the process of next step would be part of post-processing analysis, use
training models. sampling of a dataset for sanitization.
The process is limited to have performed computations
over a supervised dataset to generate the results in the form of
a Graph that can present data to check accuracy over privacy
loss.

VI. SOFTWARE DESIGN

This project follows Flask for Python-based architecture

for software design. For high-level architectural diagram is
given below in Figure 5

Fig 6: Pre-Processing Data Analysis

2. Post- Processing Data involves a complex process in the

flowchart given in Figure 7, applies the algorithm for
differential privacy using the Classification and regression
model. The final step is to train a model using the
GaussianNB mechanism to perform aggregation of the
Fig 5: Architectural diagram for Differential privacy using an dataset
algorithm

IJISRT21AUG402 www.ijisrt.com 143

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

2. The accountant module has the functionality of account

Setup to perform the data upload and preview, and the
dataset performs the masking and clustering by shuffling
columns.

Fig 7: Post-Processing Data Analysis

VII. IMPLEMENTATION

The Web application has to perform the computation on

Differential privacy open-sourced algorithm using Python-
based technologies such as Flask and other python libraries
like Scikit[13], NumPy[12], SciPy[14] in the Diffprivlib[15]. 3. Data Summarizer that displays the comparative Analysis of
The application is deployed on a localhost python-based server Non-Zero Count, Mean, Variance, and Standard Deviation,
using version 3. Code is also available on private GitHub Calculates the accuracy of the data.
account It is of a lightweight Python-based Framework called
Flask.

The below-given screenshot displays the front end and

helps to understand the navigation and deployment of the
Differential privacy library that has been integrated into open-
sourced Web Application

1. Home Page is the main page for all functional modules that
are given in the below screenshot

IJISRT21AUG402 www.ijisrt.com 144

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
A. Business use-case
Use case:
Consider use case using healthcare dataset for diabetes as
an example[16] to perform classification and regression to
generate outputs for different Epsilon content results given in
the next chapter.

1. Using the diabetes dataset, load the data with Accountant

4. Data Visualization that displays data plotted using Name and CSV File. The dataset contains columns:
Histogram dataset. A Graph plots with X-axis with Pregnancies, Glucose, blood pressure, skin thickness,
column value and Y-axis with epsilon value to compare Insulin, BMI, DiabetesPedigreeFunction, Age, Outcome
data pre-and post-processing with differential privacy on
Private and Non-Private in the below screenshot.

2. Perform Sampling and display data by masking column

that contains the private information related to a user. In
5. Train Model with three different computational this example, select BMI as confidential to mask.
mechanisms on classification and regression dataset for
GaussianNB model, Linear regression, and Logistic
Regression.

This chapter has implementation details that include a 3. In data summarizer displays the data accuracy pre- and
list of tools, commands for installation, and framework design post-application of differential privacy algorithm; those
with a screenshot of a web application; Next chapter provides values show how the accuracy is changing with each
complete in-depth details of testing business usecase. column on different epsilon values. The step helps in easy
comparison and understanding of change inaccuracy of
VIII. ANALYSIS AND RESULTS data.
Project and analysis show that the web application built
is open source using IBM's Differential Privacy Library called
Diffprivlib[15]. The Final Outcome of the Private differential
data is in the form of having values given in the form of a
graph, and simple computational values that have compared to
present the data in between private and non-Private data
results that provide graph plots to display if the Differential
privacy is applied has some amount of data leakage in this
section. For the technical analysis performed, the
computational results below Similarly for Mean Value change in comparison to the
mean of each column denotes how much data is a change in
accuracy post-processing, and further similarly for Variance

IJISRT21AUG402 www.ijisrt.com 145

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
and standard deviation values to understand the differences in
accuracy over different values of Epsilon

The above results in the graph show the values change

on an overall dataset for the different epsilon values; in every
case, if ε = ∞ chooses values randomly until it gets the perfect
fit for privacy, it will give the least data accuracy.

From the example of the diabetes dataset that performs IX. CONCLUSIONS AND FUTURE SCOPE
GaussianNB, the below results are for the epsilon values.
While Epsilon (epsilon= float("inf")) is applied with different A. Conclusion
values, check iteration The primary scope of work was to develop
computational results for a dataset using a differential privacy
1. High Epsilon (i.e., greater than 1) gives better and more algorithm from IBM's general purpose and an open-source
consistent accuracy but less privacy. Here the agreement python-based library called Diffprivlib[9] for simulation,
gave 26.56%, which depicts that data confidentiality is less experimentation, and development of this web application.
if epsilon is high. Library also includes standard tools, mechanisms, and models
with the functionality of version 0.4.1 Having implemented for
classification with GaussianNB and regression- Linear and
Logistic type of model using a supervised dataset that could
display the visualization and histogram results for the different
values of epsilon. A web application developed on packages
like Numpy[12], Scikit[13], and SciPy[14]. Improving the
efficiency of the model and provide more user-friendly data
considering personal private data.

B. Future scope of work

Roadmap for the future looks at complex data
2. Small Epsilon (i.e., less than 1) gives better privacy but  The web application has the potential of a broader scope of
worse and less consistent accuracy. Privacy is 89.58% development to include many other mechanisms and
percent while ε = ∞ provides accuracy and is the perfect fit models.
dataset.  To build an ML application using diffprivlib can be used
dynamically, not just for a private analysis.
 The library can upgrade for the latest packages and other
tools to improve the efficiency and performance of Various
datasets.
 The tool can use an unsupervised dataset that contains text
with raw format using the k-means algorithm.
 From the security standpoint, recommendations are if the
dataset is vulnerable to any attack; from IBM and Poneman
institutes data breach report estimates "Mega Breach" up to
1 to 50 million costs.[4]

Organizations adopting the emerging technologies and

securing them remotely during the pandemic will require some
The Resultant Graph for all the epsilon values versus the Machine learning of datasets to have privacy in compliance
accuracy results in the below-given screenshot with GDPR standards. The future would rely on having an
automated feature that could depend only on Artificial
intelligence and Machine Learning, and this project can have
enormous potential to develop and be made available for the
larger community and educational purposes.

IJISRT21AUG402 www.ijisrt.com 146

Volume 6, Issue 8, August – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
ACKNOWLEDGMENT [9]. N. Holohan, "Diffprivlib: Differential Privacy Library,"
2020, [Online]. Available:
I would like to acknowledge the support provided by https://diffprivlib.readthedocs.io/en/latest/index.html.
Chancellor Dr. P Shyama Raju, Dr. S.Y Kulkarni, Ex-Vice [10]. D. Su, J. Cao, N. Li, E. Bertino, and H. Jin,
Chancellor, Dr. K. Mallikharjuna Babu, Vice-Chancellor, and "Differentially private K-Means clustering," CODASPY
Dr. M. Dhanamjaya, Registrar. 2016 - Proc. 6th ACM Conf. Data Appl. Secur. Priv., pp.
26–37, 2016, doi: 10.1145/2857705.2857708.
I would like to sincerely acknowledge the great support [11]. K. S. S. Kumar and M. P. Deisenroth, "Differentially
and encouragement by Director, Corporate Training for Private Empirical Risk Minimization with Sparsity-
RACE, and all the eminent RACE Faculty who mentored with Inducing Norms," vol. 12, pp. 1069–1109, 2019,
efficacy throughout the course. [Online]. Available: http://arxiv.org/abs/1905.04873.
[12]. “Numpy’s histogram function,” [Online]. Available:
I would like to acknowledge Special thanks to all my https://numpy.org/doc/stable/reference/generated/numpy.
mentors, for helping and mentoring me throughout the Course. histogram.html.
[13]. “Scikit,” [Online]. Available: https://scikit-
And this whole course would not have been possible learn.org/stable/.
without the support of my family, my office colleagues and [14]. “Scipy,” [Online]. Available: https://www.scipy.org/.
my classmates at REVA, who made learning interesting, fun [15]. N. Holohan, S. Braghin, P. Mac Aonghusa, and K.
and collaborative. Levacher, "Diffprivlib: The IBM Differential Privacy
Library," pp. 1–5, 2019, [Online]. Available:
Last but not least, I would like to extend my heartfelt http://arxiv.org/abs/1907.02444.
thanks to the Admin, Infrastructure team, and extended [16]. "Diabetes Dataset," [Online]. Available:
support staff at RACE, and the diligent interns, for all the help https://www4.stat.ncsu.edu/~boos/var.select/diabetes.htm
and support. l.

REFERENCES

[1]. F. K. Naoise Holohan, Steve Martinelli, "IBM's open

sourced Differential Privacy library - Diffprivlib," 2019.
https://github.com/IBM/differential-privacy-library/.
[2]. A. Alnemari, R. K. Raj, C. J. Romanowski, and S.
Mishra, "Protecting Personally Identifiable Information
(PII) in Critical Infrastructure Data Using Differential
Privacy," 2019 IEEE Int. Symp. Technol. Homel. Secur.
HST 2019, pp. 6–11, 2019, doi:
10.1109/HST47167.2019.9032942.
[3]. M. Lecuyer, V. Atlidakis, R. Geambasu, D. Hsu, and S.
Jana, "Certified robustness to adversarial examples with
differential privacy," Proc. - IEEE Symp. Secur. Priv.,
vol. 2019-May, pp. 656–672, 2019, doi:
10.1109/SP.2019.00044.
[4]. IBM Security, "Cost of Data Breach Report 2020," p. 82,
2020, [Online]. Available:
https://www.ibm.com/security/digital-assets/cost-data-
breach-report/#/.
[5]. M. Abadi et al., "Deep learning with differential
privacy," Proc. ACM Conf. Comput. Commun. Secur.,
vol. 24-28-Octo, no. Ccs, pp. 308–318, 2016, doi:
10.1145/2976749.2978318.
[6]. A. Watson, "Differential Privacy towards Data Science,"
[Online]. Available:
https://towardsdatascience.com/tagged/differential-
privacy.
[7]. “De-identification Tools,” 2018, [Online]. Available:
https://www.nist.gov/itl/applied-cybersecurity/privacy-
engineering/collaboration-space/focus-areas/de-id/tools.
[8]. N. Holohan, D. J. Leith, and O. Mason, "Optimal
Differentially Private Mechanisms for Randomised
Response," IEEE Trans. Inf. Forensics Secur., vol. 12,
no. 11, pp. 2726–2735, 2017, doi:
10.1109/TIFS.2017.2718487.

IJISRT21AUG402 www.ijisrt.com 147