System, method, and program product for recognizing and rejecting fraudulent purchase attempts in e-commerce

Abstract

This disclosure relates generally to a system and method for using a machine-learning system to more accurately detect fraudulent use of credit cards on an e-commerce website and block those attempts.

Description

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH AND DEVELOPMENT

Not Applicable.

BACKGROUND

This disclosure relates generally to a system, method, and program product, for recognizing attempts to fraudulently verify credit card details and mitigating those attacks without causing unnecessary impacts to existing customers and business.

Big e-commerce companies face attacks every day in many different ways. Having a website that is open to the Internet and accepting payments over the Internet make e-commerce companies a prime target for credit card fraud. One vector of attack is based on credit card verification attacks to guess partial credit card details. This type of fraud comprises using a credit card a number of times on a website, or over multiple websites, to discover through trial-and-error details such as whether a credit card is valid, expiration date, CCV/CVV/CVV2 (card verification value) code, and other details needed to fraudulently use a credit card number, cancelling the order each time, to discover if the credit card is valid and find out specific details such as the expiration date and CCV/CVV/CVV2 number. Many websites limit the number of failed attempts that a single customer could make with a card, however many programs would not detect multiple attempts coming from different customer accounts or different IP addresses. As such, some attackers may distribute their attacks over multiple customer accounts and multiple IP addresses to avoid the software that flags a card used for multiple attempts at guessing the information. These attacks are often carried out using automated programs, so a variety of methods of detecting automated programs have evolved, but these methods can impact the customer experience, and may not prevent all distributed verification attacks. In addition, attack methods are constantly evolving, using different methods to hide a distributed verification attack from the software designed to prevent it.

What is needed is a method to mitigate against distributed credit card verification attacks spread out across different IP addresses using multiple customer accounts, while ensuring that existing regular customer traffic is not impacted. The ideal solution would be implemented by credit card companies and/or at the payment network layer, as only these groups have the ability to see multiple attempts for the same credit card that are distributed over various different web sites. However, solutions for individual companies are needed to help e-commerce companies mitigate the risk of accepting credit cards online and to help prevent distributed verification attacks from using these websites. The instant invention focuses on individual websites, being designed to prevent distributed verification attacks on a single website while ensuring that existing regular customer traffic is not impacted, and also being designed to take into account different factors and improve the outcomes over time.

The features and advantages of the disclosure will be set forth in the description that follows, and in part will be apparent from the description, or may be learned by the practice of the disclosure without undue experimentation. The features and advantages of the disclosure may be realized and obtained by means of the instruments and combinations particularly pointed out herein.

SUMMARY OF THE DISCLOSURE

A purpose of the instant disclosure is to prevent distributed verification attacks on credit cards from using an e-commerce website to obtain details about a credit card, while ensuring that regular customer traffic on the e-commerce website is not impacted. The goal is to prevent malicious users from placing orders, even if those attacks are coming from multiple different IP addresses and multiple customer accounts. Traditional fraud detection systems look at the number of orders per customer account, thus allowing those using multiple accounts to get around them. The system allows an automated method to be put into the website that can correlate IP addresses, customer identifications, and other data points that allows or denies access and employs machine learning protocols to continuously improve the system while not impacting genuine orders from customers.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a chart showing the general system of an e-commerce website with the location of the instant invention as a filtering service.

FIG. 2 is a diagram showing the machine learning feedback process of the fraud detection system.

DETAILED DESCRIPTION

For the purposes of promoting an understanding of the principles in accordance with this disclosure, reference will now be made to illustrative embodiments of the invention. It will nevertheless be understood that no limitation of the scope of the disclosure is thereby intended. Any alterations and further modifications of the inventive features illustrated herein, and any additional applications of the principles of the disclosure as illustrated herein, which would normally occur to one skilled in the relevant art and having possession of this disclosure, are to be considered within the scope of the disclosure claimed.

Before the devices, systems, processes, and methods will be disclosed and described, it is to be understood that this disclosure is not limited to the particular configurations, process steps, and materials disclosed herein, as such configurations, process steps, and materials may vary somewhat. It is also to be understood that the terminology employed herein is used for the purpose of describing particular illustrative embodiments only and is not intended to be limiting since the scope of the disclosure will be limited only by the appended claims and equivalents thereof.

In describing and claiming the subject matter of the disclosure, the following terminology will be used in accordance with the definitions set out below.

It must be noted that, as used in this specification and the appended claims, the singular forms “a”, “an”, and “the” include plural referents unless the context clearly dictates otherwise.

As used herein, the terms “comprising,” “including,” “containing,” “characterized by,” “having” and grammatical equivalents thereof are inclusive or open-ended terms that do not exclude additional, unrecited elements or method steps.

As used herein, a “bot” (short for “robot”) is an automated program that runs over the Internet. As is known in the art, some bots run automatically, while others only execute commands when they receive specific input. See https://techterms.com/definition/bot. As used herein, “artificial intelligence” or “AI” is the ability of a computer to act like a human being and/or a program designed to do so. See https://techterms.com/definition/artificial_intelligence.

The following computer systems and elements refer to systems used in the ideal embodiment of the invention described below. The invention described below is not limited to the use of the specific elements below, but as will be apparent to those skilled in the art, any similar program or platform which is equivalent may be used to effect the invention described. Nevertheless, in the interest of disclosing the preferred embodiment, specific systems will be disclosed.

As used herein, “Amazon RDS” refers to Amazon Relational Database Service, a platform used to set up, operate, and scale a relational database in the cloud. It provides cost-efficient and resizable capacity while automating time-consuming administration tasks such as hardware provisioning, database setup, patching and backups (Amazon Relational Database Service (RDS), 2018).

As used herein, “Redis” refers to an open source (BSD licensed), in-memory data structure store, used as a database, cache and message broker that has built-in replication, Lua scripting, LRU eviction, transactions and different levels of on-disk persistence, and provides high availability and automatic partitioning. (Redis, 2018)

As used herein, “Amazon ElastiCache” refers to Amazon ElastiCache for Redis, a fast, in-memory data store built an open source Redis that provides sub-millisecond latency to power internet-scale real-time applications. (Amazon ElastiCache for Redis, 2018)

As used herein, “Apache Hadoop” refers to the Apache Hadoop software library, a framework that allows for the distributed processing of large data sets across clusters of computers using simple programming models (Apache Hadoop, 2018).

As used herein, “Apache Kafka” refers to a distributed streaming platform used for building realtime streaming data pipelines that reliably get data between systems or applications and transform or react to the streams of data (Apache Kafka, 2018).

As used herein “storm cluster” refers to a cluster created using Apache Storm.

A system, method, and program product for detecting and preventing distributed verification attacks from using an e-commerce website is presented. The system prevents malicious users from placing orders on the website while allowing normal customer traffic to be unaffected. The system improves upon existing systems by detecting and preventing malicious users from using the website even when those malicious users utilize several different IP addresses and spread their use across several different customer accounts. It does this by creating an automated system that is built into an e-commerce website that correlates orders by IP address, customer ID, and other data points to allow or deny an order made on an e-commerce website. The method described is used to implement the solution in any number of websites that allow a customer to make a purchase via credit card.

One illustrative embodiment of the current invention creates an application designed to include detection of credit card verification attacks even though they may be coming in from a distributed range of IP addresses and are spread out over multiple customer accounts for an e-commerce website; identify patterns of attacks to be able to track new attempts as the attackers evolve their tactics in response to the mitigation strategies the company implements; not impact existing customer traffic; build out logic in the software application (website) to ensure only malicious users are blocked and genuine traffic is allowed without any issues; and monitor and tune the application logic over time to evolve with the changing tactics of those attempting fraud.

A system for preventing distributed verification attacks comprises a non-transitory computer readable medium containing instructions that can be integrated into an existing e-commerce website. The system may comprise a series of instructions in a non-transitory computer-readable medium. The set of instructions contained in the computer readable medium require any request to purchase from the website to be verified as not fraudulent. The instructions to complete this verification process may be stored in a computer readable medium in the form of an independent, named application, which is stored in the computer readable medium and which is run any time a user attempts to check out. In one embodiment this named fraud filtering application is called FraudFilteringService. In one embodiment this application may generate either a “blocked” or “not-blocked” response whenever it is called in response to an attempt to make a purchase. The system only allows an attempt to make a purchase if a “not-blocked” response is generated, showing that the purchaser is verified.

The independent application, FraudFilteringService, determines whether the request to make a purchase is blocked or not by using certain logic and criteria to determine whether a check-out request is likely to be fraudulent and therefore should be blocked. The logic the application uses to make this determination may include a number of factors, each of which may have a different weight. For example, this logic can include a determination based on Customer account ID, IP address, transaction amount, shipping address, products in carts, browser user agent, browser language settings, HTTP referrer, the total time spent on the website, the Frequency of the visit, the Ratio of successful orders to attempts, or the number of pages visited on the website before checkout. These factors can have different weights assigned to them and can be updated and modified through a machine learning program.

As seen in FIG. 1, the request to purchase originates in the customer's browser 101 and is sent from the customer's browser to the website 102 in part and in part is through the API-Gateway 103 of the system. In one embodiment, customer traffic is directed to an e-commerce website and is rendered by the website application 102. Subsequent calls to the website can be handled by the API-Gateway application 103, which routes requests to the required back end API interface. For example, the first request to load the checkout page goes to the website 102. However, once the page is loaded, subsequent calls, such as filling in a billing or shipping address go to the API-Gateway app 103 to be routed to the proper API interface (such as the Shipping API interface, or the AddressAPI interface) (not shown). Users of the website (whether legitimate customers or attackers), however, access only the API-Gateway application 103 and the website application 102. These are the points in the system that are connected to the filtering service 104. In addition to the filtering service 104, a cache 105 and database 106 are established.

In one illustrative embodiment the filtering service 104 takes the form of an application called “FraudFilteringService.” This application is created as a RESTful webservice to check the validity of requests coming in from the e-commerce website and the API gateway for checkout. The business logic for the accepting or rejecting the checkout request resides in this application. The filtering service evaluates if the incoming request should be blocked or not based on the data points coming in, the current business logic, and the data written previously by this application in the database. To cut down the number of calls to the database and make the solution performant, short-term caching mechanisms are implemented. This ensures that the requests are near-real-time while making sure the solution is performant.

In one illustrative embodiment, a new programming cluster is created based on Apache Kafka or a similar program to ingest the checkout requests from the website and API-Gateway, and also the final fraud decision taken for the respective checkout request. This information is stored and used in later machine-learning systems to improve the quality of the system.

In one embodiment, when a user wants to complete a purchase on an e-commerce website, the instructions in the computer readable media cause a request to be sent to the Filtering Service 104. This service sends a request to a memory data structure cache 105, which has a time to live of 2 minutes, for example. If data are not found in the cache 105, a further call is made to the database 106, and the data are then stored in the cache 105 for two minutes, for example. The cache 105 is established using a fast, in-memory data store that can provide sub-millisecond latency able to power Internet-scale real-time applications, such as Amazon ElastiCache or similar systems. The database 106 is set up as a relational database in the cloud, which can be done using Amazon RDS or a similar program.

The filtering service incorporates logic for preventing access to the website by users deemed to be performing distributed verification attacks. The system can incorporate a number of relevant criteria when analyzing a request. Some of the criteria that can help the system determine whether a distributed verification attack is taking place include the number of previous attempts to check out, the customer account ID that is checking out, the IP address from which the check out takes place, transaction amount, shipping address, what products are in the carts, the browser user agent, the browser language settings, the HTTP referrer, the total time spent on the website, the frequency of the visits to the website, the ratio of successful orders to attempts, and the number of pages visited before checkout on the website. These are examples; additional factors could be imagined that might be useful, such as credit card number or other factors. The filtering service analyzes the factors, as noted below, and then responds to the request with a “blocked” or “not-blocked” response based on the logic and the various factors that are indicative of possible fraud.

In one embodiment of the invention, when the filtering service determines that a request should be blocked, it may respond with an HTTP code 400 response (Bad Request). Another embodiment allows the filtering service to respond with an HTTP code 503 response (Internal Server Error).

Each of the relevant criteria may be given a different weight by the filtering service 104, or even no weight at all. For example, in one embodiment of the system, the system may consider only the number of previous attempts to check out, based on either the customer or the credit card number. However, another embodiment might incorporate the customer account ID and the IP address from which the purchase takes place, with appropriate weight given to each factor. The system may also incorporate machine learning processes to change the weight of each factor. As such, it will be understood that in a particular system, the specific weight given to each factor may vary with both time and with the specific website using the program. This enables the program to better respond to fraud and tailor the approach to the specific needs of the particular website on which it is being used.

FIG. 2 shows an embodiment of the system in an e-commerce environment, including a machine-learning program that adjusts the weight of each factor to determine whether to block a particular request. In this embodiment, a check-out request is made by a customer, which goes through the website and the website's API gateway 201 and 202. At the time of the request, the system calls a Checkout Action Aggregator service 203. This service is responsible for fetching aggregated data in context to the existing checkout request by referring to the available historical data stored in the database 204. The data collected are used to enrich the request data.

At the same time, the Fraud filtering service is called 205, which obtains necessary features from a Feature Store 206 for machine learning and a predicted probability score for checkout is obtained using a Model Store 207. The Fraud filtering service 205 then uses the information to create a response from the request parameters, features used, and the allowed or blocked decision, which is stored in fraud-decision-topic 209. In one illustrative embodiment the filtering service publishes the exact request parameters, features used, and response provided to fraud-decision-topic. In one embodiment, fraud-decision-topic is a Kafka topic created using Apache Kafka or similar system. Meanwhile, request logs from Website/API-gateway are channeled through the checkout requests 208 and create a log entry 210 on another Kafka topic, checkout-request-topic, which obtains request logs from the website and API gateway 201 and 202. This log entry 210 on checkout-request-topic is correlated with a log entry on fraud-decision-topic 211 using a unique UUID (universally unique identifier) generated for every request 212. Joining the logs gives the entire context of the checkout request, as well as the features used and decision taken for the checkout request.

The logs are then cleaned for invalid data, imputed for missing data, and normalized where needed. This is done using a framework that can perform fast distributed computing and allows programs to load data and query the data repeatedly, such as Apache Spark or a similar platform. These data can be used for training a machine learning model, which can be evaluated to ensure that it is performing accurately. When the model is above a threshold accuracy, it is stored in object storage and is used in the Model Store 207 for future predictions and decisions, and is then further refined by the continuing process.

In one embodiment, access logs from the website and API-Gateway are channeled through an Apache Kafka cluster are sent to a Storm cluster and the output is fed back to the Filtering Service and written to the database. This information is leveraged for all subsequent attempts and this will be the basis for the initial Machine Learning algorithm. Based on how the malicious traffic changes, the logic in the Storm cluster will be updated. A whitefish section can be provided to whitefish any customer ID in case they are wrongfully blocked.

In addition to recording and analyzing traffic patterns over time, as well as recording blocks by the fraud filtering service, the system uses machine learning logic to update and refine the fraud filtering service. In addition to the blocks made, manual corrections for given situations can be included in the machine learning process. For example, another point in the service creates a “whitefish” in the fraud filtering service to allow access by any customers who are wrongfully blocked and allows those users to make purchases on the website. The machine-learning elements of the system can be used to update and refine the systems for blocking distributed verification attacks as attackers incorporate different methods to conceal distributed verification attacks from detection software. The whitefish can be added to manually when it is determined that an order was blocked by mistake and can be used in the machine-learning process to train the system and to improve the accuracy of the filtering service 205.

In one illustrative embodiment, machine learning applications ensure that traffic patterns are recorded and analyzed over time to make the FraudFilteringService more accurate. In addition, the Checkout Action Aggregator service is responsible for fetching the aggregated data in context to the existing checkout request by referring to the available historical data.

In one illustrative embodiment, the system is implemented using AWS-EC2 cloud service and the following resources on the AWS cloud:

• • EBS Volumes for Customer Data: 4 SSD volumes with storage of 300 GB per volume, operating at 900 IOPS and having baseline throughput of 160 MBs/sec.
  • Application Servers: 2 instances running Linux on m5d.2xlarge
  • Web Servers: 2 instances running Linux on m5d.2xlarge
  • Database Servers: 2 instances running Linux on m5d.2xlargeIn this embodiment, each of the above m5d.2xlarge servers will have the following configuration: 8 vCPUs, 32 GiB Memory, 1×300 NVMe SSD drive with an I/O of up to 10 Gbps. In one illustrative embodiment, the setup includes 10 Elastic IPs per month and will have incoming data transfer limit of 20 GB/week coming in and outgoing data transfer limit of 20 GB/month. The data are backed up and stored to help with future machine learning projects.

In one illustrative embodiment, the e-commerce website and the API-Gateway run on Java. In this embodiment, the FraudFilteringService program also runs on Java and connects to the machine learning components. The preferred embodiment uses Amazon RDS as the backend relational database to store all the data from the various transactions. The cache is implemented using Amazon ElastiCache for Redis instance. The Cache may have a time to live of 2 minutes, for example, to cut down on the database calls. This time to live may be adjusted upward or downward as may be better suited for the operation of the system in a certain environment. For example, a cache with a time to live of 5 minutes, 4 minutes, 3 minutes, or one minute, or another selected time, may be instituted. Or a cache with longer or shorter term time to live may be instituted. The cache ensures that the application remains performant even when a new call is introduced for each checkout request from the e-commerce website.

Claims

1. A system for detecting and preventing distributed verification attacks on an e-commerce website having an API gateway for checkout, comprising: an e-commerce computer configured for connecting to a purchaser computer through the Internet;a non-transitory computer readable medium containing a series of fraud-detection instructions that cause a website to: run a fraud detection webservice checking the validity of requests coming in from the e-commerce website and the API gateway;wherein for each request coming in from the e-commerce website and the API gateway, the fraud detection webservice compares data about the user to a series of factors relevant to whether the purchase attempt is fraudulent and records the factors used to determine whether an attempt is fraudulent;a server connected to the Internet, wherein the server contains programming directing the system to execute the fraud-detection instructions each time a user attempts to make a purchase; andat least one machine learning algorithm for training the fraud detection system and adjusting the factors used to determine whether a distributed verification attack is taking place.

2. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include information on a number of previous attempts made.

3. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include a Customer account ID.

4. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include an IP address.

5. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include an amount of the transaction.

6. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include a shipping address.

7. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include products in carts.

8. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include a Browser user agent.

9. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include Browser language settings.

10. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include an HTTP referrer.

11. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include total time spent on the website.

12. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include frequency of visits to the website.

13. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include a ratio of successful orders to attempted orders.

14. The system of claim 1 wherein the factors used to determine whether a purchase attempt is fraudulent include a number of pages visited by the user before checkout.

15. The system of claim 1 also comprising a Checkout Action Aggregator that obtains data in context to an existing checkout request by referring to available historical data.

16. A method of detecting and preventing distributed verification attacks on an e-commerce website comprising a fraud filtering program, comprising: storing available historical data about customers and their purchases in a historical database;comparing data stored in the historical database about a user attempting to complete a purchase on a website to a series of factors relevant to whether the purchase attempt is fraudulent;using the data stored and the factors relevant to whether the purchase attempt is fraudulent to determine whether the purchase attempt is fraudulent;recording the factors used to determine whether an attempt is fraudulent;causing the website to execute the fraud filtering program each time a user attempts to make a purchase;preventing the purchase from being completed if the attempt is deemed to be fraudulent;sending the information on the factors used to determine whether an attempt is fraudulent to at least one machine learning algorithm to train the fraud filtering program and adjust the weights of factors used to determine whether an attempt is fraudulent;using the recorded factors to train a system through machine-learning to better stop fraudulent attempts to use credit cards; andincorporating the newly trained system into the fraud filtering program and adjusting the weight of the factors to determine whether an attempt is fraudulent in consequence of a distributed verification attack taking place.

17. The method of claim 16 wherein the factors used to determine whether a purchase attempt is fraudulent include information on a number of previous attempts made.

18. The method of claim 16 wherein the factors used to determine whether a purchase attempt is fraudulent include a Customer account ID.

19. The method of claim 16 wherein the factors used to determine whether a purchase attempt is fraudulent include an IP address.

20. The method of claim 16 wherein the factors used to determine whether a purchase attempt is fraudulent include an amount of the transaction, a shipping address, products in carts, a Browser user agent, Browser language settings, an HTTP referrer, total time spent on the website, frequency of visits to the website, a ratio of successful orders to attempted orders, or the number of pages visited by the user before checkout.