Adaptive ML Thresholds for Credit Card Enumeration Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models used for credit card validation face inaccuracies due to static thresholds, leading to excessive validation fees during enumeration attacks, as they fail to distinguish between legitimate and illegitimate charges effectively.
Innovation Solution
Implementing a dynamic threshold adjustment mechanism using a second machine learning model to determine whether a merchant is under an enumeration attack, allowing the first model to adapt its threshold based on the second model's output, thereby improving prediction reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a static threshold is used in the machine learning model, then the model structure is simple and easy to implement, but the model fails to accurately distinguish between legitimate and illegitimate charges during enumeration attacks
Solution Approach 1:
The patent implements a dynamic threshold mechanism where the classification threshold is no longer fixed but adapts based on the predicted probability distribution of the input data. The system calculates an optimal threshold dynamically for each prediction based on the learned probability distribution, allowing the model to adjust its decision boundary according to the specific characteristics of each validation request, thereby improving reliability without requiring a fundamentally more complex model architecture
Solution Approach 2:
The patent changes the parameter of the threshold from a static constant to a dynamic value that is derived from the probability distribution of the input features. By computing the threshold based on the predicted probabilities and the specific characteristics of each validation request (such as merchant history, transaction patterns, and risk indicators), the system optimizes the threshold parameter adaptively, improving prediction accuracy while maintaining reasonable model complexity
2Reliability
If validations are submitted for all charges to ensure legitimate transactions are processed, then legitimate charges are not blocked, but excessive validation fees are incurred during enumeration attacks
Solution Approach 1:
The system dynamically adjusts the classification threshold parameter based on the predicted probability of illegitimate charges. When the model detects patterns consistent with enumeration attacks (such as multiple rapid validations from the same merchant or unusual transaction patterns), it adjusts the threshold to be more stringent, blocking more validations to reduce fees. For normal traffic, the threshold remains more permissive to ensure legitimate charges are processed, thus optimizing the balance between validation accuracy and fee reduction
Solution Approach 2:
The system incorporates feedback mechanisms where the outcomes of validations (whether they were legitimate or fraudulent, and whether they were blocked or processed) are used to refine the probability distribution model. This feedback loop allows the system to learn from past decisions and improve its threshold adjustment strategy over time, reducing validation fees by increasingly accurately identifying which validations can be safely blocked while maintaining high reliability for legitimate transactions
Data Source
AI summary
In an example embodiment, a solution is provided wherein a threshold used by a classifier in a first machine learning model is dynamically set by a second machine learning model. More particularly, the threshold may be selected from two or more different threshold settings, based on the output of the second machine learning model. This acts to improve the reliability of predictions made by the first machine learning model in certain use cases where circumstances not adequately captured by the first machine learning model can affect the accuracy of the threshold used by the first machine learning model.