LLM Fine-Tuning for Tabular Fraud Prediction With Privacy Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLMs) struggle to effectively process tabular data due to the need for text-based input, lacking a method to convert tabular data into LLM-friendly formats while preserving data privacy and enabling pattern learning from both internal and external sources.
Innovation Solution
A system and method that converts tabular data into narrative prompts, uses embeddings to represent learned features, and fine-tunes LLMs with human feedback to enhance predictive capabilities, employing techniques like LoRA and reinforcement learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If tabular data is converted into narrative prompts for LLM processing, then LLM capability utilization is improved, but data privacy protection is worsened due to external sharing requirements
Solution Approach 1:
The patent introduces an intermediary conversion layer that transforms tabular data into narrative prompts without requiring raw data to leave the secure environment. The conversion happens within the sandboxed LLM execution context, allowing LLM processing capability to be utilized while maintaining data privacy through isolation boundaries.
Solution Approach 2:
Instead of sharing actual sensitive data externally, the system creates a copy of the data in a transformed narrative format within the LLM's execution environment. This copy can be processed by the LLM without exposing the original sensitive tabular data to external systems, thus maintaining privacy while enabling capability utilization.
2Object-affected harmful factors
If traditional ML models are used for tabular data analysis, then data privacy is maintained, but model development time and complexity increase
Solution Approach 1:
The patent makes the LLM universal by enabling it to handle tabular data through narrative conversion, allowing the same model to perform multiple functions including fraud detection, pattern recognition, and data analysis. This multi-functionality eliminates the need for separate specialized models, reducing development time while maintaining privacy through sandboxed execution.
Solution Approach 2:
The system changes the input parameter format from raw tabular data to narrative prompts, allowing LLMs to process structured data in a format they naturally understand. This parameter transformation enables faster development by leveraging pre-trained LLM capabilities rather than training new models from scratch, while the sandboxed environment maintains data privacy.
3Measurement precision
If LLMs are fine-tuned with human feedback for tabular data, then prediction accuracy is improved, but computational resources and training complexity increase
Solution Approach 1:
The patent performs preliminary conversion of tabular data into narrative prompts before LLM processing. This pre-processing step simplifies the training complexity by providing LLMs with data already in their preferred format, reducing the need for complex training procedures while maintaining high prediction accuracy through human feedback fine-tuning.
Data Source
AI summary
A method for training a machine learning (ML) model using a large language model (LLM) is provided. A system for detecting fraud which utilizes the LLM-trained ML model trained is also provided. An artificial intelligence (AI)-based method for monitoring alerts is also provided. The method for training an ML model using an LLM includes receiving tabular data for training the ML model, generating one or more natural-language strings comprising information from the tabular data, generating, via a base LLM, one or more prompts and completions based on the one or more generated natural-language strings, pre-training the base LLM using a plurality of generated prompts and completions, updating the base LLM via supervised learning using a cross-entropy loss function with ground-truth labels, and fine-tuning the updated LLM via reinforcement learning with human feedback using a reward model and a proximal policy optimization model to produce the LLM-trained ML model.