VIP Call Girls Service Hitech City Hyderabad Call +91-8250192130
Short-Story-Writing 255.pptx
1. Credit Card Fraud Detection
Using Neural Network Auto
Encoders
Raparla Sravani
016656601
San Jose State University
2. How Credit Cards Is Frauded?
Types Of Fraud Detections:
➢ Usually credit card frauds are committed with the intention of quickly
profiting off the hard-earned cash of unwitting or otherwise technologically
unsavvy victims.
➢ Lack of proper security standards, the use of stolen credit cards,
impersonating the actual account holder, phishing sites impersonating real
transaction portals, and other factors are among the main causes of credit card
fraud.
➢ Fraudulent Uses Of Credit Cards:
1. Card Related Frauds
2.Merchant Related Frauds
3.Internet Frauds
3. A.Card Related Frauds
● Application fraud occurs when a fraudster substitutes the real
application by identifying their personal information.
● When a customer loses their debit or credit card, they assume it
was simply misplaced, but the fraudster is actually withdrawing
funds from the victim's account.
● (Card Not Present): A transaction can be completed even if the
card is not physically in your hands. Even with knowledge of the
card's expiration date and number, fraudsters can still manage to
perpetrate crimes.
4. B. Merchant Related Frauds C. Internet Frauds
● arises when vendors give
criminals access to customers'
private information for
financial benefit.
● They have elaborate strategies
to commit fraud by luring users
to use certain items, such
shopping carts, by allowing
them to enter their personal
information.
● Comparable to merchant-
related frauds in several
ways.
● Every keystroke the user
makes is recorded by this
spyware, which is then sent to
a fraudster who utilizes it to
commit transaction-related
fraud.
● Camera Scam on a Cell
Phone
6. ● Encoder and Decoder are the two components of the
model I have used
● Instead of forecasting new target values, the goal is to
reconstruct the inputs as target values.
● Although there are connections within each layer of an
autoencoder network, there are none between the layers.
● The model's cost function is a straightforward mean
squared error loss, which has the following
representation:
● 1/𝑛 S (𝑖 = 1) (𝑦𝑖 − 𝑦 𝑖) 2
● Our model evaluation would be significantly stricter and
less prone to random errors thanks to this configuration.
7. V. EXPERIMENTAL SETUP
Data Collection
Dataset Wrangling
Data Visualization & Analysis
Data Preprocessing
● One hot encoding
● Z Score Normalization
● Data Clipping
Data Modelling
8. Data Collection:
● We get a fake dataset from Kaggle for credit card fraud
detection.
● Six million financial transactions total, broken down into
five groups: CASH IN, CASH OUT, DEBIT, PAYMENT,
and TRANSFER.
● Importing pandas causes it to run slowly.
● We therefore have a different framework called "DASK."
● Out of 6362620 transactions in the dataset, there were 8213
fraudulent transactions, or roughly 0.12% of all
transactions.
9. Dataset Wrangling:
● We must make sure that the dataset is clean and adequate for
training in order to create an effective model.
● We first accomplish so by comprehending the dataset
through visuals and analysis.
● Later, by expanding on this knowledge, the dataset can be
cleaned and prepped as needed.
10. Data Visualization & Analysis:
● Our transaction dataset contains 10 columns in all. The data
columns contain a combination of boolean, string, and
numeric values. The Table indicates the name, data type,
and description of each column. 1.
● There are 8213 fraudulent transactions in the dataset,
which has a total of 6362620 transactions.
● Prior to the training of any model, this significant class
imbalance would need to be corrected.
13. Data Preprocessing:
● The normalization and other steps we take during data
preprocessing are based on the analysis we completed in the
previous stage. At this stage, preparatory techniques like Z Score
normalization and One Hot encoding are used.
● The most common technique for handling categorical data when the
number of categories is particularly limited is one hot encoding.
● In particular, Z Score Normalization aids in preventing the outlier
problems that might arise with Min-Max Normalization.
● Data clipping: This method is used when the non-outlier values
must fall inside a certain range.
14. Raw dataset to Processed dataset looks as shown in Fig.4 and Fig.5
respectively:
15.
16. Data Modelling:
● Based on how the training proceeds we can suitably alter the training to help
the model learn better
17. CONCLUSION & FUTURE WORK:
● Detecting credit card fraud is a very significant issue.
● In order to address the issue of credit card fraud, we suggested an
autoencoder model design.
● In the future, we'll work to increase the computational efficiency of our
model and train and test it on a larger dataset.
● Additionally, we'll aim to include more intriguing elements like
transaction frequency and geographical area.
18. REFERENCES:
● Philip ,Chan & Fan, Wei & Prodromidis, Andreas & Stolfo, Salvatore. (2019). Credit
Card Fraud Detection Distributed Data Mining. IEEE Intelligent Systems. 14.
10.1109/5254.87786570
● Sam Maes, et al., ―Using bayesian and neural networks credit cards fraud detection,‖ p.
4, August 2002.
● https://medium.com/@sravani.raparla/credit-card-fraud-detection-using-neural-network-
auto-encoders-9adbed0bf29f