THE ROLE OF

XAI AND FL
IN FINANCIAL FRAUD DETECTION
Transparent, Trustworthy, and Collaborative
Solutions
Abstract:
■ Traditional fraud detection systems often rely on centralized Machine Learning (ML) models, which
pose risks to data privacy and security. To address these challenges, this project introduces a novel
privacy-preserving fraud detection framework using Federated Learning (FL) and
Convolutional Neural Networks (CNN). Federated Learning allows multiple financial institutions to
collaboratively train a global fraud detection model while ensuring that sensitive customer data
remains decentralized and secure. The CNN algorithm is employed for its ability to effectively learn
spatial patterns and detect complex relationships in transaction data, thereby improving model
accuracy.

■ To enhance transparency and interpretability, Explainable AI (XAI) techniques, such as SHAP

(SHapley Additive exPlanations), are integrated into the framework. XAI ensures that the CNN model's
predictions are understandable, allowing human experts to trust and validate decisions. Experimental
evaluations demonstrate that the proposed FL-based CNN model achieves high accuracy, precision,
and recall in detecting fraudulent transactions across federated datasets. This research highlights the
potential of combining Federated Learning with deep learning algorithms and interpretability
techniques to create robust, transparent, and privacy-preserving fraud detection systems.
Turning the Tables: Biased, Imbalanced, Dynamic
Tabular Datasets for ML Evaluation
■ Key Points
1. BAF Dataset Suite:
1. First privacy-preserving, large-scale, tabular dataset for ML.
2. Focuses on fraud detection in bank account openings.
2. Challenges Addressed:
1. Bias: Group size, prevalence, and separability disparity.
2. Imbalance: Fraud prevalence is 0.85%–1.5%.
3. Temporal Dynamics: Shifting distributions over time.
■ Relevance to Previous Work
• Provides realistic, dynamic datasets to evaluate fairness and
performance of ML models.
• Addresses shortcomings of older static tabular datasets like UCI
Adult and COMPAS.
Credit Card Fraud Detection Using Deep Learning
Based on Auto-Encoder
■ Key Points
1. Objective:
1. Develop a deep learning model using Auto-Encoders (AE) for
credit card fraud detection.
2. Model & Approach:
1. Multilayer Auto-Encoder with "tanh" and "ReLU" activation
functions.
2. Evaluated on Accuracy, Precision, Recall, F1 Score, and AUC-
ROC.
■ Relevance to Previous Work
• Advances existing ML-based fraud detection techniques by:
• Addressing imbalanced datasets.
• Using unsupervised learning with Auto-Encoders for better
anomaly detection.
Fraud Detection Using Machine Learning and Deep
Learning
■ Key Points
1. Objective:
1. Benchmark multiple Machine Learning (ML) and Deep Learning
(DL) methods for fraud detection.
2. Techniques Used:
1.ML: k-Nearest Neighbor (KNN), Random Forest, Support Vector
Machines (SVM).
2.DL: Autoencoders, Restricted Boltzmann Machine (RBM), Deep
Belief Networks (DBN), Convolutional Neural Networks (CNN).
■ Relevance to Previous Work
• Demonstrates that ensemble methods and CNN can achieve better
performance on unbalanced datasets for fraud detection.
Towards Federated Learning at Scale: System
Design
■ Key Points:
1. Concept: Federated Learning (FL) trains models on decentralized
data, ensuring privacy and data locality.
2. Architecture: A scalable FL system using TensorFlow was deployed,
supporting synchronous model updates with Secure Aggregation.
3. Applications: Implemented in mobile devices for use cases like
Gboard keyboard predictions and item ranking.
■ Relevance to Previous Work:
1. Both papers emphasize privacy-preserving machine learning
techniques: The current work's secure aggregation parallels the
fraud detection document's Federated Learning (FL) for private,
decentralized training.
2. Transparency: Explainable AI (XAI) from the fraud detection
document complements the privacy protocols in this paper, ensuring
model interpretability in sensitive applications.
Privacy Intrusiveness in Financial-Banking Fraud
Detection
■ Key Points:
1. Focus: Analyzes privacy intrusiveness in fraud detection algorithms
from a regulatory and consumer perspective.
2. Privacy Concerns: Explores GDPR-compliant fraud detection
methods and the balance between privacy and anti-fraud efficiency.
3. Algorithms:
1. Static rules and behavior analysis for fraud detection.
2. Data minimization principles vs. invasive data collection (device
details, transaction profiles).
■ Relevance to Previous Work:
1. Federated Learning (FL): Aligns with privacy by design
principles from the previous document, emphasizing decentralized
approaches to mitigate privacy concerns.
2. Explainable AI (XAI): Complements the need for transparency in
fraud detection algorithms as emphasized in both documents.
Algorithm:
■ While the traditional centralized learning model requires sensitive data to be
shared with a central server, FL allows collaborative training of a global model
across decentralized datasets, ensuring data privacy.
■ Key Components of the Algorithm:
1. Federated Learning (FL):
1. Objective: Train a global model collaboratively across multiple financial
institutions without sharing raw data.
2. Federated Averaging Algorithm: Each participating client (financial
institution) trains a local model using its own data and computes local
updates (weight changes).
1. The updates are aggregated centrally to refine the global model.
2. The global model is then sent back to all clients for further training.
Continues...
■ Explainable AI (XAI):
• Objective: Improve model transparency by making predictions
interpretable for human experts.
• SHAP (SHapley Additive exPlanations):
• SHAP values attribute the contribution of each feature towards the
model's predictions.
• Features with high SHAP values are identified as significant
contributors to the detection of fraud.
Methodology:
■ Data Preprocessing:
1. Handling Missing Values:
Numerical attributes are imputed with the mean, and categorical
attributes are Imputed with the mode.
2. Data Balancing : The Synthetic Minority Over-Sampling
Technique (SMOTE) is used to address the class imbalance
problem.
Continues:
■ CNN Model Design (Proposed Modification):
• Instead of a standard Deep Neural Network (DNN), a Convolutional
Neural Network (CNN) is proposed to process financial transaction
data.
• The CNN will leverage convolutional layers to extract patterns from
transaction features, improving detection accuracy.
• The architecture may include:
• Convolutional Layers for pattern extraction.
• Dense (Fully Connected) Layers for classification.
• Activation Functions: ReLU (Rectified Linear Unit) for hidden
layers and Sigmoid for the output layer.
Continues:

Federated Learning Process:

■ Each financial institution (client) trains a local CNN model on its private data.
■ The model updates (not raw data) are sent to a central server for aggregation.
■ The global CNN model is updated iteratively and redistributed to clients.

Integration of XAI:

■ After training the CNN model, SHAP values are calculated to explain which features most influence
predictions.
■ Visualizations of SHAP plots help human experts understand the model’s decisions, ensuring
transparency and trust.
Data Set Used:
■ Dataset Attributes
1. Transaction-Specific Features:
1.Transaction Amount: Value of the transaction.
2.Transaction Time: Time elapsed since the first transaction.
2. Customer Behavior Features:
1.Credit Score: Customer’s creditworthiness.
2.Account Balance: Available account balance.
3.Transaction Frequency: Number of transactions per day/week.
4.Location: Geographical location of the transaction.
Continues:
1. Device and Channel Features:
1.IP Address: Source of the transaction.
2.Device ID: Unique identifier of the device used.
3.Payment Mode: Card type (e.g., Visa, MasterCard).
2. Derived Features:
1.Transaction Pattern: Deviation from typical spending behavior.
2.Merchant Category: Type of store where the transaction
occurred.
■ Highlights:
• Combines transactional, behavioral, and risk-based features.
• Ensures rich data inputs for detecting anomalies and fraud patterns.
Conclusion:
■ Federated Learning (FL):
• Enables collaborative training of fraud detection models across
institutions.
• Ensures data privacy by keeping sensitive data decentralized.
■ Explainable AI (XAI):
• Provides transparency and builds trust through explainable predictions.
• Techniques like SHAP highlight feature contributions in decision-making.
■ Key Achievements:
• Improved fraud detection accuracy.
• Ensured privacy preservation and model fairness.
• Adaptability to dynamic fraud patterns.
■ Impact:
Transparent, Trustworthy, and Privacy-Preserving Fraud Detection for
modern financial systems.