Explainable AI (XAI) in Deep Learning Models for Credit Card Fraud Detection

Documentation: Report Documentation.pdf
Main analysis notebook: XAI_methods.ipynb
Run the live main notebook:
Models: architectures

Description

Credit card fraud represents an important and growing challenge in the financial sector, causing significant monetary losses worldwide. According to the data from the UK Finance, fraudsters stole over £1.3 billion in 2021 alone through authorised and unauthorised fraud, with card fraud accounting for a significant part of these damages ¹. Detection of fraud transactions presents several important challenges that increase the need for advanced computational approaches.

The project utilized and evaluated the XAI methods on Deep Learning architectures that are widely known to be used in detecting credit card transaction fraud, including CNN and LSTM with attention mechanisms, trained on Sparkov's synthetic dataset. The main contribution lies in the integration and comparative analysis of three Explainable AI methods: SHAP, LIME and Anchors. Research further evaluates the effectiveness of each XAI method based on Faithfulness, Monotonicity, and Completeness metrics.

Academic Poster
(Click to view PDF)

Key Features:

Multi-Model Comparison:
Assess and compare different deep learning architectures for fraud detection.
Explainable AI Integration:
- SHAP: Interactive visualisations (force plots, summary plots, dependence plots) that elucidate feature contributions.
- LIME: Local interpretable explanations (in progress) to detail individual predictions.
- Anchors: Rule-based explanations offering high-precision insights.
Robust Evaluation:
- Visual tools to explore feature impacts locally, supporting transparent decision-making.
- Extensive testing on synthetic data with XAI performance metrics to validate both model interpretability.

System Architecture

The fraud detection system using XAI comprises several integrated modules:

Data Collection & Model Design:
Gathers synthetic transaction data (Sparkov dataset) and design model architectures (CNN, LSTM). Data is loaded and versioned for reproducibility.
Data Preprocessing:
Raw transaction data is cleaned, normalised, and balanced using the SMOTE method.
Model Training:
Separate pipelines are implemented for training the CNN and LSTM models.
XAI Integration:
Post-training, XAI techniques are applied to generate explanations for model predictions.
Explainability Performance Evaluation:
Comprehensive XAI evaluation metrics, including Faithfulness, Monotonicity, and Completeness are computed.

---
config:
  theme: neo
---
flowchart TD
 subgraph Data_Preprocessing["<b>Data Preprocessing</b>"]
        A["Raw Transaction Data"]
        B["Feature Engineering"]
        C["Preprocessed Data"]
        D["Data Normalization"]
        E@{ label: "Class Balancing (<span style=\"--tw-scale-x:\"><b>SMOTE</b></span>)" }
  end
 subgraph Model_Training["<b>Model Training</b>"]
        F1["<b>CNN Pipeline</b>"]
        F2["<b>LSTM Pipeline</b>"]
        F1a["Feature Extraction"]
        F1b["CNN Model Training"]
        F1c["Model Optimization"]
        M1["Trained CNN Model"]
        F2a["Sequence Preparation"]
        F2b["LSTM Model Training"]
        F2c["Model Optimization"]
        M2["Trained LSTM Model"]
  end
 subgraph XAI_Integration["<b>XAI Integration</b>"]
        X1a["Apply <b>SHAP</b> Explanations"]
        X1b["Apply <b>LIME</b> Explanations"]
        X1c["Apply <b>Anchors</b> Explanations"]
        X3["Feature Importance Analysis"]
        X4["Local Explanation Generation"]
  end
 subgraph Performance_Evaluation["<b>XAI Metrics Evaluation</b>"]
        G1["<b>Faithfulness</b>"]
        G2["<b>Monotonicity</b>"]
        G3["<b>Completeness</b>"]
        Z["Visualisation"]
        Z1["XAI Evaluation Report"]
  end
    n2@{ label: "<b style=\"color:\">Sparkov</b> <span style=\"color:\">Dataset Acquisition</span>" } --> A
    A --> B
    B --> C
    C --> D
    D --> E
    E --> F1 & F2
    F1 --> F1a
    F1a --> F1b
    F1b --> F1c
    F1c --> M1
    F2 --> F2a
    F2a --> F2b
    F2b --> F2c
    F2c --> M2
    M1 --> X1a & X1b & X1c
    X1a --> X3 & X4
    X1b --> X3 & X4
    X1c --> X3 & X4
    G1 --> Z
    G2 --> Z
    G3 --> Z
    Z --> Z1
    M2 --> X1a & X1c & X1b
    Z1 --> n1["<b>Deployment Decision</b>"]
    X4 --> n1
    X3 --> G1 & G2 & G3
    D@{ shape: rect}
    E@{ shape: rect}
    n2@{ shape: rect}
     A:::preprocessing
     B:::preprocessing
     C:::preprocessing
     D:::preprocessing
     E:::preprocessing
     F1:::training
     F2:::training
     F1a:::training
     F1b:::training
     F1c:::training
     M1:::training
     F2a:::training
     F2b:::training
     F2c:::training
     M2:::training
     X1a:::xai
     X1b:::xai
     X1c:::xai
     X3:::xai
     X4:::xai
     G1:::evaluation
     G2:::evaluation
     G3:::evaluation
     Z:::evaluation
     Z1:::evaluation
     n2:::data
     n1:::decision
    classDef data fill:#b3e6ff,stroke:#35424a,stroke-width:3px,color:#222222
    classDef preprocessing fill:#ffcccb,stroke:#35424a,stroke-width:2px,color:#222222
    classDef training fill:#c4e17f,stroke:#35424a,stroke-width:2px,color:#222222
    classDef xai fill:#f3d1ff,stroke:#35424a,stroke-width:2px,color:#222222
    classDef evaluation fill:#ffeb99,stroke:#35424a,stroke-width:2px,color:#222222
    classDef decision fill:#f7c6c7,stroke:#35424a,stroke-width:3px,color:#222222
    classDef subgraphLabel font-size:18px, stroke-width:5px, font-weight:bold
    class Data_Preprocessing,Model_Training,XAI_Integration,Performance_Evaluation subgraphLabel
    
    linkStyle 0 stroke:#ff6b6b,stroke-width:2px,fill:none
    linkStyle 1 stroke:#ff6b6b,stroke-width:2px,fill:none
    linkStyle 2 stroke:#ff6b6b,stroke-width:2px,fill:none
    linkStyle 3 stroke:#ff6b6b,stroke-width:2px,fill:none
    linkStyle 4 stroke:#ff6b6b,stroke-width:2px,fill:none
    linkStyle 5 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 6 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 7 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 8 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 9 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 10 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 11 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 12 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 13 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 14 stroke:#77DD77,stroke-width:2px,fill:none
    linkStyle 15 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 16 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 17 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 18 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 19 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 20 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 21 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 22 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 23 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 24 stroke:#ffcc00,stroke-width:2px,fill:none
    linkStyle 25 stroke:#ffcc00,stroke-width:2px,fill:none
    linkStyle 26 stroke:#ffcc00,stroke-width:2px,fill:none
    linkStyle 27 stroke:#ffcc00,stroke-width:2px,fill:none
    linkStyle 28 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 29 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 30 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 31 stroke:#ff9797,stroke-width:3px,fill:none
    linkStyle 32 stroke:#ff9797,stroke-width:3px,fill:none
    linkStyle 33 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 34 stroke:#d277ff,stroke-width:2px,fill:none
    linkStyle 35 stroke:#d277ff,stroke-width:2px,fill:none

System Architecture Diagram
Can't see the diagram? View system architecture image

Data Sources

Synthetic Data Generation: To generate synthetic credit card transaction data, including fraudulent transactions, refer to the Sparkov Data Generation repository.
Combined Dataset: The combined dataset from Sparkov Data Generation, converted into a standard format, can be accessed here on Kaggle.

Key Visualizations

SHAP Feature Importance Analysis

SHAP values showing global feature importance across the dataset

SHAP Waterfall Plot showing how each feature contributes to CNN's prediction for transaction #1044

SHAP Force Plot displaying the cumulative effect of all features on CNN's prediction for transaction #1044

Local Explanation Example

LIME explaining an individual of LSTM's fraud prediction for transaction #1044

Rule-Based Insights

Anchors generating interpretable rules of LSTM's fraud prediction for transaction #2025

XAI Methods Explainability Evaluation

Summary Performance Results

Method	Faithfulness	Monotonicity	Completeness
SHAP	0.602	0.447	0.171
LIME	0.325	0.467	0.217
Anchors	0.364	0.478	0.028

Confidence-Based Performance Analysis

Method	Very Low	Low	Borderline	High	Very High
Faithfulness
SHAP	0.629	0.594	0.620	0.544	0.602
LIME	0.312	0.304	0.317	0.326	0.378
Anchors	0.407	0.348	0.393	0.285	0.353
Completeness
SHAP	0.089	0.139	0.106	0.265	0.322
LIME	0.289	0.129	0.127	0.195	0.283
Anchors	0.012	0.029	0.000	0.048	0.059
Monotonicity
SHAP	0.513	0.429	0.421	0.422	0.389
LIME	0.536	0.471	0.464	0.419	0.386
Anchors	0.547	0.507	0.411	0.461	0.400

Model-Specific XAI Performance

Metric		CNN			LSTM
	SHAP	LIME	Anchors	SHAP	LIME	Anchors
Faithfulness	0.443	0.254	0.315	0.761	0.396	0.412
Monotonicity	0.464	0.488	0.486	0.429	0.445	0.469
Completeness	0.208	0.295	0.033	0.134	0.139	0.022

Key Findings

Explainability Analysis:
- SHAP demonstrated the highest faithfulness (0.602), providing the most reliable feature attributions.
- LIME showed balanced performance across metrics with good completeness (0.217).
- Anchors excelled in monotonicity (0.478) but with limited coverage (0.028).
Feature Importance: Transaction amount, merchant category, and transaction hour emerged as the most influential features across both models.
Confidence Analysis: XAI methods showed varying performance across prediction confidence levels, with SHAP maintaining the most consistent performance (0.544-0.629 faithfulness) across all confidence bins.

This repository is a valuable resource for researchers and practitioners interested in the intersection of fraud detection and explainable AI , providing a foundation for further exploration and development in this critical area of financial security.

Project Directory Hierarchy

Project Root/
├── XAI_methods.ipynb ............................ Explainable AI methods implementation  
├── Siddhartha_CNN.ipynb ......................... CNN model implementation  
├── Ibtissam_LSTM.ipynb .......................... LSTM model implementation  
├── utils.py ..................................... Utility functions  
├── requirements.txt ............................. Dependencies  

├── architectures/ ............................... Trained model storage  
│   ├── model_1_Siddhartha_CNN_acc99/  
│   └── model_2_Ibtissam_LSTM_acc98/  
├── data/ ........................................ Data files and results  
│   ├── predictions.csv  
│   ├── stratified_samples.csv  
│   └── xai_metrics.json  
├── visualisation/ ............................... Visualisation outputs  
├── README.md .................................... Project brief  
├── Report Documentation.pdf ..................... Project Report Documentation
├── presentation.pdf ............................. Project Presentation
├── poster.pdf ................................... Academic Poster
└── .gitignore ................................... Git configuration

Acknowledgments

I would like to express my sincere gratitude to my supervisor Dr. Hyunkook Lee at the University of Huddersfield, for his invaluable guidance, expertise, encouragement to explore, and patient feedback thoughout this work.

I also want to thank the researchers for their brilliant previous XAI research, which made this project possible.

UK Finance. (2022). Annual Fraud Report 2022. https://www.ukfinance.org.uk/policy-and-guidance/reports-and-publications/annual-fraud-report-2022 ↩

Navigation Menu

Repository files navigation

Explainable AI (XAI) in Deep Learning Models for Credit Card Fraud Detection

Description

System Architecture

Data Sources