US-12619920-B2 - Distributed adaptive machine learning training for interaction exposure detection and prevention

Abstract

Embodiments of the present invention provide for a distributed adaptive learning transaction fraud detection and prevention system has a meta-model system that accesses a fraud meta-model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; receives from at least one sub-system a sub-system best performing fraud model; updates the fraud meta-model based at least in part on the sub-system best performing fraud model; and transmits the updated fraud meta-model to the at least one sub-system; and at least one sub-system receives the updated fraud meta-model transmitted from the meta-model system; accessing a sub-system fraud model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; and updates the sub-system fraud model with the updated fraud meta-model transmitted from the meta-model system.

Inventors

• Ioannis Giokas

Assignees

• Verge Capital Limited

Dates

Publication Date

20260505

Application Date

20220729

Claims (18)

1. 1 . A method of training a machine-learning (ML) algorithm to detect and prevent exposure, the algorithm trained by a meta-model system comprising a memory device, the memory device comprising computer readable instructions and a processing device operatively coupled with the memory device for executing the computer readable instructions to cause the processing device to: (i) access from the memory device an exposure meta-model comprising a real-time exposure detection and prevention engine and an interaction database comprising interaction exposure decision data and interaction exposure feedback data; (ii) receive from at least one sub-system a sub-system best performing exposure model; (iii) update the exposure meta-model in response to feedback data via a feedback loop from the at least one sub-system that is non-collocated with the meta-model system associated with the exposure meta-model transmitting its best performing exposure model trained on local transaction data to the meta-model system, and by using a plurality of machine learning techniques to ensemble a plurality of classification models, thereby resulting in a plurality of ensembling results and subsequently stacking the plurality of ensembling results in order to increase accuracy; and (iv) transmit the updated exposure meta-model to the at least one sub-system, wherein the at least one sub-system is communicatively coupled to and non-collocated with the meta-model system and comprises a sub-system memory device comprising computer readable instructions and a sub-system processing device operatively coupled with the sub-system memory device for executing the computer readable instructions to cause the sub-system processing device to: (i) receive the updated exposure meta-model transmitted from the meta-model system; (ii) access a sub-system exposure model comprising a real-time exposure detection and prevention engine and a interaction database comprising interaction exposure decision data and interaction exposure feedback data; and (iii) update, in real-time, the sub-system exposure model with the updated exposure meta-model transmitted from the meta-model system.
2. 2 . The method of claim 1 , further comprising: (a) receiving, by the at least one sub-system, new transaction data; (b) applying, by the at least one sub-system, the real-time exposure detection and prevention engine to the new interaction data, thereby resulting in a new interaction exposure decision; (c) communicating, by the at least one sub-system, the new interaction exposure decision to the interaction database; and (d) communicating, by the at least one sub-system, new interaction exposure feedback data to the interaction database.
3. 3 . The method of claim 2 , further comprising: (e) training, by the at least one sub-system, the interaction exposure decision data and interaction exposure feedback data of the interaction database using a plurality of machine learning techniques, thereby resulting in trained data; and (f) updating, by the at least one sub-system, the real-time exposure detection and prevention engine based at least in part on the trained data.
4. 4 . The method of claim 3 , further comprising: (g) identifying, by the at least one sub-system, a best performing exposure model; and (h) transmitting, by the at least one sub-system, the best performing exposure model to the meta-model system.
5. 5 . The method of claim 1 , wherein using the plurality of machine learning techniques comprises ensembling a plurality of classification models, thereby resulting in a plurality of ensembling results and subsequently stacking the plurality of ensembling results in order to increase accuracy.
6. 6 . The method of claim 5 , wherein the plurality of classification models comprises at least one selected from the group consisting of a gradient boosting model, a random forest model, an isolation forest model, an isolation forest model alongside a multi-layer neural network model, and/or an isolation forest model alongside a genetic algorithm.
7. 7 . A distributed adaptive learning transaction fraud detection and prevention system comprising: (a) a meta-model system comprising a memory device comprising computer readable instructions and a processing device operatively coupled with the memory device for executing the computer readable instructions to cause the processing device to: access from the memory device a fraud meta-model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; receive from at least one sub-system a sub-system best performing fraud model; update the fraud meta-model in response to feedback data via a feedback loop from the at least one sub-system that is non-collocated with the meta-model system associated with the fraud meta-model transmitting its best performing exposure model trained on local transaction data to the meta-model system; and transmit the updated fraud meta-model to the at least one sub-system; and (b) at least one sub-system communicatively coupled and non-collocated with the meta-model system comprising a memory device comprising computer readable instructions and a processing device operatively coupled with the memory device for executing the computer readable instructions to cause the processing device to: (i) receive the updated fraud meta-model transmitted from the meta-model system; (ii) access a sub-system fraud model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; and (iii) update, in real-time, the sub-system fraud model with the updated fraud meta-model transmitted from the meta-model system.
8. 8 . The system of claim 7 , wherein the at least one sub-system has a processing device for executing the computer readable instructions further to cause the processing device to: (a) receive new transaction data; (b) apply the real-time fraud detection and prevention engine to the new transaction data, thereby resulting in a new transaction fraud decision; (c) communicate the new transaction fraud decision to the transaction database; and (d) communicate new transaction fraud feedback data to the transaction database.
9. 9 . The system of claim 8 , wherein the at least one sub-system has a processing device for executing the computer readable instructions further to cause the processing device to: (a) train the transaction fraud decision data and transaction fraud feedback data of the transaction database using a plurality of machine learning techniques, thereby resulting in trained data; and (b) update the real-time fraud detection and prevention engine based at least in part on the trained data.
10. 10 . The system of claim 9 , wherein the at least one sub-system has a processing device for executing the computer readable instructions further to cause the processing device to: (a) identify a best performing fraud model; and (b) transmit the best performing fraud model to the meta-model system.
11. 11 . The system of claim 7 , wherein using the plurality of machine learning techniques comprises ensembling a plurality of classification models, thereby resulting in a plurality of ensembling results and subsequently stacking the plurality of ensembling results in order to increase accuracy.
12. 12 . The system of claim 11 , wherein the plurality of classification models comprises at least one selected from the group consisting of a gradient boosting model, a random forest model, an isolation forest model, an isolation forest model alongside a multi-layer neural network model, and/or an isolation forest model alongside a genetic algorithm.
13. 13 . A method for distributed adaptive learning transaction fraud detection and prevention using a meta-model system comprising a memory device comprising computer readable instructions and a processing device operatively coupled with the memory device for executing the computer readable instructions to cause the processing device to perform a set of actions and a sub-system operatively coupled with the meta-model system comprising a memory device comprising computer readable instructions and a processing device operatively coupled with the memory device for executing the computer readable instructions to cause the processing device to perform a second set of actions, the method comprising: (a) accessing, by the meta-model system from the memory device, a fraud meta-model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; (b) receiving, by the meta-model system and from at least one sub-system, a sub-system best performing fraud model; (c) updating, by the meta-model system, the fraud meta-model in response to feedback data via a feedback loop from the at least one sub-system that is non-collocated with the meta-model system associated with the fraud meta-model transmitting its best performing exposure model trained on local transaction data to the meta-model system; (d) transmitting, by the meta-model system, the updated fraud meta-model to the at least one sub-system; (e) receiving, by the at least one sub-system, the updated fraud meta-model transmitted from the meta-model system; (f) accessing, by the at least one sub-system, a sub-system fraud model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; and (g) updating, by the at least one sub-system in real-time, the sub-system fraud model with the updated fraud meta-model transmitted from the meta-model system.
14. 14 . The method of claim 13 , further comprising: (i) receiving, by the at least one sub-system, new transaction data; (j) applying, by the at least one sub-system, the real-time fraud detection and prevention engine to the new transaction data, thereby resulting in a new transaction fraud decision; (k) communicating, by the at least one sub-system, the new transaction fraud decision to the transaction database; and (l) communicating, by the at least one sub-system, new transaction fraud feedback data to the transaction database.
15. 15 . The method of claim 14 , further comprising: (m) training, by the at least one sub-system, the transaction fraud decision data and transaction fraud feedback data of the transaction database using a plurality of machine learning techniques, thereby resulting in trained data; and (n) updating, by the at least one sub-system, the real-time fraud detection and prevention engine based at least in part on the trained data.
16. 16 . The method of claim 15 , further comprising: (o) identifying, by the at least one sub-system, a best performing fraud model; and (p) transmitting, by the at least one sub-system, the best performing fraud model to the meta-model system.
17. 17 . The method of claim 13 , wherein using the plurality of machine learning techniques comprises ensembling a plurality of classification models, thereby resulting in a plurality of ensembling results and subsequently stacking the plurality of ensembling results in order to increase accuracy.
18. 18 . The method of claim 17 , wherein the plurality of classification models comprises at least one selected from the group consisting of a gradient boosting model, a random forest model, an isolation forest model, an isolation forest model alongside a multi-layer neural network model, and/or an isolation forest model alongside a genetic algorithm.

Description

CROSS-REFERENCE TO RELATED APPLICATION This application claims the benefit of priority of U.S. provisional patent application No. 63/228,403, titled “DISTRIBUTED ADAPTIVE LEARNING TRANSACTION FRAUD DETECTION AND PREVENTION SYSTEM,” filed on Aug. 2, 2021, which is incorporated herein in its entirety by this reference. FIELD This invention relates generally to the field of fraud detection and prevention, and more particularly embodiments of the invention relate to a distributed adaptive learning transaction fraud detection and prevention system. BACKGROUND Financial institutions, especially card issuers and card processors, are facing a major problem related to unauthorized card operations by fraudsters & criminals. This is further accelerated based on the changes in consumer behavior that are expanding the attack surface exponentially, like the use of multiple channels (e.g. online, mobile, phone, in-store etc.) to transact with their cards. On top of that, consumers are oversharing on social media & paying little attention to details (e.g. fraudulent sites, phishing emails etc.) & hope between networks (e.g. Wi-Fi, 3G, 4G, LTE, 5G etc.) making the life of fraud detection systems & experts very difficult. This leads to a very challenging environment to balance between customer experience vs. costs (e.g. HR, tools, financial losses, insurance policies etc.) vs. fraud rate where financial institutions have to continuously optimize the rule-based systems so as to code each fraud scenario in an evolving landscape, which requires a lot of time, resources & expertise that is hard to get & expensive (either in the form of experts or automation tools); while in parallel such system generate a number of alerts creating noise for analysts & frustration for users. This has an impact both to the reputation and the bottom line of financial institutions since consumers do not like their transactions to be declined or being victims of fraud. The main difficulty is the highly unbalanced data. There are various solutions for the imbalance problem such as the application of oversampling or under-sampling techniques. These techniques do not provide sufficient performances on real-world data, and the development of per financial institution model is not solving the imbalanced and skewed data problem in the long-term. BRIEF SUMMARY Embodiments of the present invention address the above needs and/or achieve other advantages by providing apparatuses and methods that provide a distributed adaptive learning transaction fraud detection and prevention system. Within this context we have developed a real-time streaming card fraud detection & prevention system. A system that is continuously self-learning & fine-tuning its models; that enable it to have increased accuracy & reduced false positives. This is based on a novel approach toward machine learning model design and deployment architecture of the overall system. Machine Learning-based methods can continuously improve the accuracy of fraud prevention solutions according to information about each cardholder's behavior. These AI solutions are suited perfectly not only for credit cards but can be implemented for e-commerce fraud detection purposes, as well as many other industries were financial transactions are involved. However, fraud detection requires a substantial amount of planning before using machine learning algorithms at it, whereas a common theme that persists is continued human supervision to make the solutions more efficient and effective. Human analysts are still needed to investigate fraudulent patterns. Practically speaking, eliminating fraud wholly might not be possible, but our system aims to optimize the resources required to better tackle it. In fraud analysis, efficiency remains an important issue that should be focused on to achieve a high fraud detection rate. Efficiency guarantees the performance of the fraud detection models, even an increase of 1% accuracy rate is beneficial because it will have an advantageous impact in detecting fraudulent activities and fraudsters. The goal of our system is to be a fraud analytics framework that deals with the yet unsolved class imbalance problem to enhance the efficiency of the fraud analytics systems. Embodiments of the invention a distributed adaptive learning transaction fraud detection and prevention system comprising a meta-model system comprising a memory device comprising computer readable instructions and a processing device operatively coupled with the memory device for executing the computer readable instructions to cause the processing device to access from the memory device a fraud meta-model comprising a real-time fraud detection and prevention engine and a transaction database comprising transaction fraud decision data and transaction fraud feedback data; receive from at least one sub-system a sub-system best performing fraud model; update the fraud meta-model based at least in part on the sub-system best per