AN ADAPTIVE PREDICTIVE FINANCIAL FRAUD DETECTION APPROACH USING DEEP LEARNING METHODS ON A BIG DATA PLATFORM

Abstract

Fraud, waste, and abuse in many financial systems are estimated to result in significant losses annually. Predictive analytics offer government and private financial institutions the opportunity to identify, prevent or recover such losses. This work proposed a novel Big Data driven approach for fraud detection based on Deep Learning methods. A supervised Deep Learning solution leveraging Big Data was shown to be an effective Fraud predictor. Additionally, an unsupervised method based on anomaly detection using deep auto encoders was proposed for when there is few or no labelled data. The two methods presented offered adaptive and predictive Fraud detection through improved Analytics. Future work will look into how the two methods can be integrated into an effective tool for enhanced Fraud detection.

Chapter 1

Introduction

1.1 Background to the Study

Fraud refers to the intentional illegal exploitation of a system which results in injury of an oblivious entity. Financial fraud involves the exploitation of financial systems which results in the loss of financial resource, the most prominent being monetary although other damages such as loss of integrity are possible. Fraud, waste, and abuse in many financial systems are estimated to result in significant losses annually running into billions of US dollars. Furthermore, the proliferation of the internet has exposed financial systems to diverse fraudsters using different mechanisms to exploit financial systems. This provided an explode in attack patterns which rendered the once effective case-based fraud detection solutions no more effective as the computational complexity increases with each new detected fraud. More seriously, their is a higher tendency for first-time frauds going undetected. The case-based detection methods are also slow as a successful exploit could multiply if the solution took time to be integrated into the system. This problem can only be addressed with an online (on-the-y) adaptive (able to detect new frauds) solution. Also of concern to financial fraud detector solutions is, the prediction strength that indicates a Fraud detector's ability to correctly identify both known and novel Frauds. This is usually a direct function of how much fraud samples there are to model a solution. The emergence of Big Data and its Analytics has provided financial fraud detection experts with verse amount of data that will enhance the detection models. Such solutions that use Big Data to model offer more comprehensive solutions. A complete fraud detection model thus, must have the following properties:

1. Adaptive: This refers to the following abilities:

·         Ability to detect fraudulent activities within a short period of time. This is also referred to as its alertness.

·         Ability to detect first-time fraudulent activities with high accuracy.

2. Predictive: This refers to the following abilities:

·         Ability to detect all new instances of fraudulent activities that have happened in the past. This is very difficult to achieve if there is no data with a considerable description of previous transactions.

Over the years solutions have been proposed to provide effective solutions to financial frauds. Most of the models proposed to address the Fraud detection model property 1 have been statistical models that try to detect outliers in the data set (See [27], [21] and [8]). This follows after the assumption that fraudulent transactions will behave abnormally different from legitimate transactions. An abnormal pattern of behavior (i.e. an Outlie ) is fagged\suspicious." More recent, Machine Learning methods have been used to develop more effective models (See [4], [11], [7] and [9]). The emergence of Big Data analytic tools provided means to address Fraud detection model property 2. Such technology allows the integration of data from various sources used to model and predict financial fraud. For example location data of a fraudster, social-media activity and credit card information can be reconciled to trace a fraudulent transaction to him. However, Big Data analytics brings with it challenges that limit application of techniques used to address fraud detection model property

1. Such challenges are enumerated in [24], some of which are:

1. High-dimensionality and data reduction,

2. Data quality and validation,

3. Data cleansing,

4. Feature engineering,

5. Data representations and distributed data sources,

6. Data sampling

Much research has gone into addressing some of the above issues so that existing models that work for \small" data can scale-up to work with Big Data, for example [16], [26] proposed improvements that address high-dimensionality, others are [13], [29]. However all these attempts might not have scaled well as they were not originally designed to handle Big Data complexity. Deep learning is one technique that has the capability to handle such complex abstractions. It is good at analyzing and organizing large amount of unsupervised data. Most raw data in Big Data Analytics are largely unlabeled and uncategorised, which are ideally suited for Deep learning algorithms.

Department: Computer Science (M.Sc Thesis)

Format: MS Word

Chapters: 1 - 5, Preliminary Pages, Abstract, References, Appendix.

No. of Pages: 46

Price: 20,000 NGN

In Stock

Buy Now

Our Customers are Happy!!!

See What Some of Our Customers are Saying About Us