Black-Box Adversarial Sample Generation for Fraud Detection Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing black-box attack techniques for generating adversarial examples on tabular data are limited by high query counts and lack applicability, posing a significant risk to fraud detection models in banking and other industries.
Innovation Solution
A server system employs a gradient-free optimization method using encoders initialized with random weights, iteratively optimizing them to generate adversarial samples through an evolution-based approach, reducing query counts and enhancing the robustness of fraud detection models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing black-box attack techniques are used to generate adversarial examples on tabular data, then adversarial samples can be produced, but the query count becomes excessively high
Solution Approach 1:
The patent pre-trains an encoder model on the tabular dataset before the actual adversarial example generation process. This preliminary encoding step creates a compressed representation of the data that can be rapidly manipulated to generate adversarial examples, significantly reducing the number of queries needed compared to direct optimization methods
Solution Approach 2:
The encoder acts as an intermediary between the original tabular data and the adversarial example generation process. By transforming the data through the encoder first, the system creates a intermediate representation that is more efficient to optimize, thereby reducing the computational queries required for generating effective adversarial examples
2Reliability
If existing black-box attack techniques are applied to fraud detection models, then adversarial samples can be generated, but the applicability and effectiveness are limited
Solution Approach 1:
The patent changes the parameter representation by encoding tabular data into a compressed latent space using the pre-trained encoder. This parameter transformation allows existing image-based adversarial techniques to be effectively applied to tabular fraud detection data, improving both applicability and effectiveness while maintaining model security
Solution Approach 2:
The patent adapts and applies image-based adversarial attack methodologies to the tabular fraud detection domain by using the encoder to create a comparable representation structure. This copying of successful techniques from one domain (images) to another (tabular data) enhances the versatility and effectiveness of the attack methods
Data Source
AI summary
Various embodiments relate to methods and systems for generating adversarial samples. The method performed by a server system includes accessing a set of payment transaction samples from transaction database. The method includes initializing a plurality of encoders, weights of each of the plurality of encoders being randomly initialized. Further, the method includes computing a set of initial adversarial samples using the plurality of encoders based on the set of payment transaction samples. Further, the method includes optimizing the plurality of encoders to generate a plurality of evolved encoders. Further, method includes computing a plurality of fitness scores for the plurality of evolved encoders. Further, the method includes determining a top evolved encoder from the plurality of evolved encoders based on the plurality of fitness scores. Further, the method includes generating a set of final adversarial samples using the top evolved encoder based on the set of payment transaction samples.