To clarify your question, I am assuming you mean the application of machine learning in "credit/debit card fraud detection" and "bank fraud detection". This problem is essentially an anomaly detection problem in machine learning. This is due to the following characteristics:

• The number of positive cases (fraudulent attempts) are small
• The application involves deviation from the normal pattern

It is just the matter of obtaining features that would represent the normal pattern of usage of a legitimate user and deviation from this observed pattern would infer the possibility of a fraudulent attempt. 

The following links should make this clear.

• Presentation on Anomaly Detection, by Jing Gao
• Fraud and Anomaly Detection Using Oracle Advanced Analytic Option 12c

http://www.cse.buffalo.edu/~jing/cse601/fa12/materials/outlier_detection.pdf

http://www.oracle.com/technetwork/database/options/advanced-analytics/fraudoaa12c-1979769.pdf

http://www-users.cs.umn.edu/~banerjee/papers/09/anomaly.pdf

The general steps could be;

1. Enlist the attributes of a nonfradulent system and fradulent system.

2. Generate training set. With two kinds ( fradulent : 1, nonfradulent : 0 ) or diffrent kind ( like no fraud, medium fraud, high fraud etc.. with fuzzy gradation)

3. Train a network such as SOFM or RBFN etc... as it is a classification problem.

Check this paper: 

http://link.springer.com/chapter/10.1007%2F978-3-642-33362-0_25

Agree to Jayaran answer.

I suggest you to collect the real data of bank/credit card fraud cases. Use these data to construct training set. And then define fraud categories as label of that training (High, Medium, etc.). Check cluster analysis to know the pattern of fraud. of Maybe you can refer to Ebenezer answer to construct training set. At least to know the attributes that can be used.

Training set can be used to evaluated the best algorithm (classification algorithm) to detect fraud based on defined labels. Use confusion matrix for accuracy measurements. 

Thank you all

use of statistcial concept SVM/ELM can be utilized to perform classification using ML

Suggestions:

http://www.ulb.ac.be/di/map/adalpozz/pdf/Dalpozzolo2015PhD.pdf

https://www.kaggle.com/mlg-ulb/creditcardfraud

http://mlg.ulb.ac.be/BruFence

Conference Paper An Assessment of Streaming Active Learning Strategies for Re...

Article SCARFF : a Scalable Framework for Streaming Credit Card Frau...

Article Learned lessons in credit card fraud detection from a practi...

Article Credit Card Fraud Detection: A Realistic Modeling and a Nove...

Article Streaming Active Learning Strategies for Real-Life Credit Ca...

Working Paper A Machine Learning Approach to Improve the Efficiency of Fak...