Fraud Detection System using Deep Neural Networks

My presentation slide in Re-Work Deep Learning Summit at Singapore, 2017

1. Fraud Detection System using Deep Neural
Networks
Hendri Karisma

2. Fraudulent Transaction
Payment Fraud (phising, account
take over, carding)
System abuse (promo, content,
account, logistic and payment
mothodsespecially COD)
Fraud not only result in financial losses
but also produce some reputational
risk.
Some security measures taken by
bank or another multinational
finance service.
[E. Dumanet al, 2013]

3. State of the art
•There are several research in fraud detection
area using some methods :
–GASS 82.78%-91% and MBO algorithm 91.3%-
94.35%
–ANN 91.74%
–SVM 83.06%
[E. Dumanet al, 2013]
–Copula-based method, extreme outlier
elimination, PCA, naïve bayes, regression logistic,
k-nearest neighbors, etc.

4. Annual Reports Cybersource

5. Annual Reports Cybersource

6. State of the art
BS (bivariate statistics)
Feature
Extractions
PCA (principal component
analysis)
Information Gain
PCA + IG = GPCA
Etc.

7. Why Deep Learning
•Dataset is high nonlinearity
•Amount of data
•A lot of features
•Mostly unlabeled data

8. Deep neural networks

9. Standard neural networks

10. Standard neural networks

11. Backpropagation

12. Deep neural networks
[karismaetal,2016]

13. Pre-learning
•Denoisingautoencoder
•Restricted boltzmannmachine

14. Autoencoder

15. Pre-training1 2 3

16. Deep neural network for bank repricinggap forcasting
•Equals network topology
•High nonlinearity
•Almost all attributes have continuous values
•Using autoencoder
•Minimum mean squared error : 10-9
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1 31 61 91 121 151 181 211 241 271 301
MSE (10^-4)
Iteration (10^2)
SB DNN

17. Network topologies
[karismaet al, 2016]

18. Algorithm Parameters
•Minimum mean squared error : 10^-8
•Learning rate : 0.75
•Momentum : 0.5
•Topologinetwork : equal
•Hidden Layer : 3 Hidden Layer
•Neuron/Hidden Layer : (26, 26, 26)
•Activation function : sigmoid function
•Autoencoder(pre-training) parameters :
–Minimum squared error : 10-5
–Max epoch: 2000
–Learning rate : 0.5
–Momentum : 0.75
–Activation function : sigmoid function

19. Dataset
•Dataset : 4000
•Fraud : 32 (confirm fraud)
•Good transaction : 2000
•False address cases : 2157
•Suspect transaction : 500
•Attributes : +/-102
•Non-linearity : High

20. Feature engineering
•Card Verification number (for BIN number)
•Postal address
•Google maps lookup (distance between shipping
and billing)
•Telephone area lookup
•Credit history
•Customer order history
•Order velocitymonitorng
•IP Geolocation
•Value Similarity (shipping and billing address,
customer email and customer name)

21. Feature engineering (Velocity)
•Mask card number given : billzip, custip, email, name,
shipzip. (just count)
•Mask card number given : billzip, custip, email, name,
shipzip. (changing)
•Email given : billzip, custip, name, masked number, shipzip
(just count
•Email given : billzip, custip, name, masked number, shipzip
(changing)
•Then billzip, custip, name, and shipzip.
•so on…
Then compute the gradient.

22. Feature engineering
It will add more than 60 features to dataset.
•Standard value
•Look-up value
•Velocity Value
0
2
4
6
8
10
0 2 4 6 8
count
email given card changing
change card
Linear (change card)
0
0.5
1
1.5
2
2.5
0 2 4 6 8
count
email given chard changing
change card
Linear (change card)
0
1
2
3
4
5
6
0 5 10 15 20 25 30
count
email count of transaction

23. Result
•Accuracy : 89.467
•ConvutionMatrix
•MSE : 8.31 x 10^-6
Fraud PassClass/origin
1221 89 Fraud
69 121 Pass
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
iterasi255075100125150175200225250275
MSE (10^
-
3)
Iteration(10^2)
Deep Neural Network for FDS

24. Challenges
•Unbalancing dataset
•Fraud is transaction perspective to fraud is person
perspective
•Event data (from checking page, order until
transaction/checkout)
•GPU optimization
•Network model architecture
•Social network features (text and network)
•Restricted Boltzmann machine & another pre-
training
•Graph theory

25. THANK YOU