LLM40FD

This is the official implementation of LLM40FD: Unlocking the Potential of LLM for Anonymous Zero-shot Fraud Detection.

Abstract: Credit card fraud detection within the realm of financial security faces challenges such as data imbalance, large-scale anonymized transaction datasets, and the need for system-specific model training. Past methods often fail to address these aforementioned issues simultaneously. Utilizing a single model results in a lack of zero-shot capability without adaptation for real-world scenarios. This paper introduces LLM40FD, a novel framework that leverages a Large Language Model (LLM) to overcome these obstacles in anonymous zero-shot fraud detection. LLM40FD addresses the aforementioned challenges in credit card fraud detection by employing a Distribution-based One-Class function and the Walking Embedding, without reliance on labeled data or fine-tuning in downstream. Additionally, LLM40FD enhances the model’s ability to detect fraudulent patterns and define robust decision boundaries. This is achieved through a dual-augmentation strategy and implicit contrastive learning, which generate enriched positive and negative samples. Our experiments demonstrate that LLM40FD not only achieves state-of-the-art performance in the full-shot setting but also exhibits strong zero-shot capability even with limited training data. Furthermore, we conduct additional experiments to validate the effectiveness and working mechanism of LLM40FD.

Architecture

Paradigms for Credit Card Fraud Detection

The framework of LLM40FD.

Performance Comparison

Quick start

TODO

Datasets

TODO. Refer to our paper.

Package

TODO.

Citation

@ARTICLE{zhijie_LLM40FD,
  author={Yang, Kaixiang and Zhong, Zhijie and Sun, Song and Yu, Zhiwen and Chen, C. L. Philip and Zhang, Tong},
  journal={IEEE Transactions on Computational Social Systems}, 
  title={LLM40FD: Unlocking the Potential of LLM for Anonymous Zero-Shot Fraud Detection}, 
  year={2025},
  volume={12},
  number={6},
  pages={4606-4619},
  keywords={Fraud;Data models;Training;Feature extraction;Zero shot learning;Training data;Legged locomotion;Adaptation models;Time series analysis;Sun;Data augmentation;fraud detection;large language model (LLM);zero-shot},
  doi={10.1109/TCSS.2025.3563954}}

TODO List

TODO

Argument Parser Documentation

This document provides a detailed explanation of the command-line arguments used in the LLM40FD project. These arguments are parsed using the argparse module in Python.

Basic Configuration

Argument	Type	Default	Choices	Description
--seed	int	2024	N/A	Random seed for reproducibility.
--debug	int	0	N/A	Debug mode. A value of 1 enables debug mode, while 0 disables it.
--saving_name	str	'normal'	N/A	Name of the directory for saving results.
--zero_shot	int	1	N/A	Enable zero-shot learning. A value of 1 enables it, while 0 disables it.

Data Loader

Argument	Type	Default	Choices	Description
--dataset1	str	'shu'	['20k', 'credit', 'jop', 'shu']	Type of the first dataset.
--dataset2	str	"20k,credit,jop,shu"	N/A	Type of the second dataset. The string should be a comma-separated list of dataset names.
--root_path	str	r'/home/xxx/anomaly_data/fraud_detection/' tip! replace it to /your/dataset/path	N/A	Root path of the data files.

Model Definition

Argument	Type	Default	Choices	Description
--enc_in	int	7	N/A	Encoder input size. Note that this is not necessary as the encoder input size can be obtained automatically.
--d_model	int	128	N/A	Dimension of the model.
--strides	list	[1, 3, 5, 11]	N/A	Strides of the walking embedding.
--llm_layers	int	3	N/A	Number of layers in the large language model (LLM).
--llm_model	str	'GPT2'	['GPT2', 'LLAMA', 'None']	Type of the large language model to use.
--llm_dim	int	768	[4096, 768]	Dimension of the large language model. For example, LLama7b has a dimension of 4096, GPT2-small has a dimension of 768, and BERT-base has a dimension of 768.
--device	str	'cuda'	N/A	Device to use for training, either 'cuda' for GPU or 'cpu' for CPU.

Optimization

Argument	Type	Default	Choices	Description
--num_workers	int	10	N/A	Number of workers for the data loader.
--train_epochs, -ep	int	1	N/A	Number of training epochs.
--few_data_per	float	0.1	[0.05, 0.1, 0.3, 0.5, 1.]	Percentage of data to use for few-shot learning.
--batch_size, -bs	int	64	N/A	Batch size of the training input data.
--learning_rate, -lr	float	0.0001	N/A	Learning rate of the optimizer.
--loss	str	'OC_CE'	N/A	Loss function to use for training.
--pos_r	float	0.8	N/A	Positive sampling ratio.
--neg_r	float	2	N/A	Negative sampling ratio.