PaperlayerView exact revision on arXiv

Source-linked AI summary

Transparency and Privacy: The Role of Explainable AI and Federated Learning in Financial Fraud Detection

Tomisin Awosika, Raj Mani Shukla, Bernardi Pranggono

arXiv:2312.13334v1cs.LGcs.AIcs.CR

TL;DR

Bank fraud detection must contend with rare fraudulent transactions, imbalanced datasets, privacy constraints, and a need for transparent decisions. This paper proposes combining FL and XAI for collaborative fraud detection without directly sharing customer data. The reported FL model accuracy is 93%.

  • Problem

    Fraud is less frequent than legitimate activity in many transaction datasets, and privacy constraints prevent sharing customer data for centralized training.

    Source: §I · p. 1

  • Method

    The paper combines FL, which keeps data local while institutions collaborate through model updates, with XAI for transparent model decisions.

    Source: §I · pp. 1–2

  • Results

    The FL model reaches an accuracy score of 93%.

    Source: B. Model Performance Metrics · p. 7

  • Takeaways & Limitations

    The paper presents FL as a privacy-preserving way for financial institutions to collaboratively train fraud-detection models without directly sharing customer data.

    Source: §I · p. 1

Abstract

from arXiv · show

Fraudulent transactions and how to detect them remain a significant problem for financial institutions around the world. The need for advanced fraud detection systems to safeguard assets and maintain customer trust is paramount for financial institutions, but some factors make the development of effective and efficient fraud detection systems a challenge. One of such factors is the fact that fraudulent transactions are rare and that many transaction datasets are imbalanced; that is, there are fewer significant samples of fraudulent transactions than legitimate ones. This data imbalance can affect the performance or reliability of the fraud detection model. Moreover, due to the data privacy laws that all financial institutions are subject to follow, sharing customer data to facilitate a higher-performing centralized model is impossible. Furthermore, the fraud detection technique should be transparent so that it does not affect the user experience. Hence, this research introduces a novel approach using Federated Learning (FL) and Explainable AI (XAI) to address these challenges. FL enables financial institutions to collaboratively train a model to detect fraudulent transactions without directly sharing customer data, thereby preserving data privacy and confidentiality. Meanwhile, the integration of XAI ensures that the predictions made by the model can be understood and interpreted by human experts, adding a layer of transparency and trust to the system. Experimental results, based on realistic transaction datasets, reveal that the FL-based fraud detection system consistently demonstrates high performance metrics. This study grounds FL's potential as an effective and privacy-preserving tool in the fight against fraud.

I. INTRODUCTION

Bank fraud detection is challenged by imbalanced, confidential transaction data and diverse fraud patterns across banks. The study proposes combining federated learning (FL) and explainable AI (XAI) to preserve privacy while making model decisions interpretable.

I. INTRODUCTION

  • The introduction frames bank account fraud as a serious threat to the global economy, institutional trust, and individuals’ financial well-being.It cites the UK Finance 2022 report on annual losses from fraud.Source: §I · p. 1
  • Fraudulent transactions are less frequent than legitimate ones, while confidential data and differing fraud patterns complicate centralized model training.The introduction identifies these characteristics as obstacles to robust fraud detection.Source: §I · p. 1
  • FL trains models across institutions by sharing aggregated updates rather than customer data, supporting collaboration while keeping data local.The introduction contrasts this multi-bank collaboration with centralized learning confined to one bank.Source: §I · pp. 1–2
  • The proposed system integrates XAI with FL so model decisions can be interpreted alongside privacy-preserving collaborative training.The introduction presents transparency and trust as reasons for adding XAI to the FL-based approach.Source: §I · p. 2
  • The listed contributions also include a DNN for recognizing fraud patterns across federated databases and a web-based application to demonstrate practicability.The contribution list identifies both the tailored model and the application deployment.Source: §I · p. 2

II. RELATED WORK

Related work surveys fraud’s changing forms, machine- and deep-learning approaches, and the privacy challenges of centralized training. It motivates an FL architecture for banking fraud detection that incorporates XAI and is deployed through a web-based framework.

II. RELATED WORK

  • The review describes fraud as evolving with technological change, including newer forms such as mobile telecommunications scams and computer breaches.It also summarizes the fraud triangle’s pressure, opportunity, and rationalization, with a later extension adding the act, concealment, and conversion.Source: §II · p. 2
  • Prior studies examine KNN, SVM, ensemble classifiers, Random Forest, AdaBoost, Auto-Encoders, and RBMs for fraud detection, including highly imbalanced data and evolving fraud.The review also describes Auto-Encoder reconstruction errors and RBMs’ pattern learning as anomaly-detection mechanisms.Source: §II · p. 2
  • Centralized fraud-detection training can require transferring confidential data, raising latency, security, and privacy concerns and facing legal or owner-imposed limits on sharing.The review contrasts that approach with decentralized learning.Source: §II · p. 2
  • FL keeps raw data local while sending model updates for server aggregation, and is described as adaptable to IID and non-IID data.The review connects this approach to privacy constraints on data sharing between banks.Source: §II · p. 3
  • The paper proposes an FL architecture for banking fraud detection, integrates XAI for transparency and trust, and deploys the platform using a web-based framework.The related-work section presents this as its response to the problem.Source: §II · p. 3

III. METHODOLOGY

The methodology uses a server-coordinated FL architecture in which banks train local DNNs and share model updates rather than raw data. It combines this privacy-preserving training with XAI for interpretable decisions.

A. System architecture

  • Each institution trains a local DNN on proprietary data, then periodically sends model insights to a coordinating server to refine a global model without moving the data.The architecture diagram is described as showing the structure and flow of this FL approach.Source: A. System architecture · p. 3, Figure 1 · p. 3
  • The server initializes and distributes a global model, aggregates returned client updates, and synchronizes the refined model for subsequent training rounds.The passage says aggregation typically involves averaging weights and may include validation against a held-out set.Source: A. System architecture · pp. 3–4
  • Clients train the received model locally, send updates rather than raw data, and use the shared model for fraud detection; the system also integrates XAI for interpretable decisions.The proposed model is a fine-tuned DNN designed to recognize fraud patterns across federated databases.Source: A. System architecture · pp. 3–4

B. Proposed Federated Learning-based model

The proposed federated-learning approach replaces centralized training with client-side updates aggregated into a global model, while keeping raw data local. Federated averaging combines local updates across clients and repeats the process over multiple rounds.

B. Proposed Federated Learning-based model

  • Clients compute model updates locally, and the server aggregates them to update the global model rather than performing the update centrally.The approach modifies standard SGD by computing several updates on each client before server-side averaging.Source: B. Proposed Federated Learning-based… · p. 4
  • Federated averaging forms the global update as a weighted sum of local updates, with weights based on each client’s number of data points.The passages define n_k as the number of data points on client k and n as the total across clients.Source: B. Proposed Federated Learning-based… · p. 4
  • The process repeats for several rounds until convergence, communicating model updates rather than raw data to help maintain data privacy.The paper presents federated averaging as a compromise between local and centralized learning that draws on diverse local data sources.Source: B. Proposed Federated Learning-based… · p. 4

C. Explainable AI integration

The paper presents interpretability as necessary in finance because model decisions have real-world implications, and describes XAI as a way to provide transparency. It highlights SHAP as a feature-importance method and notes that feature importance can aid debugging and refinement.

C. Explainable AI integration

  • In finance, model decisions have real-world implications, and XAI techniques integrated with FL can provide stakeholders insights into those decisions.The paper links those insights to enhanced confidence in model decisions.Source: C. Explainable AI integration · p. 4
  • Feature importance can aid model debugging and refinement by addressing potential pitfalls or biases.Source: C. Explainable AI integration · p. 5
  • SHAP attributes the difference between an instance’s prediction and the average prediction to features, using each feature’s average contribution across possible feature combinations.The paper describes SHAP as a unified measure of feature importance.Source: C. Explainable AI integration · p. 5

IV. IMPLEMENTATION DETAILS

The implementation section outlines preliminary checks, data processing, federated-learning development, and integration of explainable-AI techniques.

IV. IMPLEMENTATION DETAILS

  • The implementation begins with preliminary checks.Source: §IV · p. 5
  • Data processing is one of the implementation steps.Source: §IV · p. 5
  • The implementation includes federated-learning development and integration of explainable-AI techniques.Source: §IV · p. 5

A. Dataset

The proposed fraud dataset contains 29,042 records and 32 features, and its class imbalance is addressed with SMOTE.

A. Dataset

  • The dataset contains 29,042 entries across 32 features, including integer, floating-point, and categorical or string columns.It has 17 integer, 10 floating-point, and 5 categorical or string columns.Source: A. Dataset · p. 5

B. Data balancing

  • SMOTE addresses the highly imbalanced class distribution by generating synthetic feature-space samples until the two classes have equal counts.The dataset is described as highly imbalanced before balancing.Source: B. Data balancing · pp. 5–6

C. Data Pre-Processing

Preprocessing imputes missing values and removes outliers, while feature transformations prepare inputs for a three-layer neural network trained on an 80/20 split.

C. Data Pre-Processing

  • Missing numerical values are imputed with column means, categorical values with modes, and floating-point outliers beyond 1.5 × IQR from Q1 or Q3 are excluded.A correlation heatmap depicts pairwise linear relationships among numeric attributes and can reveal strong correlations or potential multicollinearity.Source: C. Data Pre-Processing · p. 6

D. Feature Selection

  • Income is converted into ten discrete bins, while selected categorical features are one-hot encoded into binary columns for model input.The encoded features include employment status, housing status, payment type, source, and device operating system.Source: D. Feature Selection · p. 6

E. Deep Learning model

  • The binary classifier is a three-layer dense network with 64- and 32-neuron ReLU layers and a single sigmoid output, optimized with Adam and binary cross-entropy.Source: E. Deep Learning model · p. 6

F. Training and Validation

  • The preprocessed data is split into 80% training and 20% testing, and the training set is divided into three parts for federated learning.Source: F. Training and Validation · p. 7

G. Performance Metrics

Model performance is assessed with several binary-classification metrics, and SHAP is used to interpret feature contributions relative to a baseline.

G. Performance Metrics

  • Accuracy, precision, recall, and F1-score are used to assess binary classification, with F1 balancing precision and recall for imbalanced classes.The evaluation also describes confusion-matrix outcomes for true positives, false positives, false negatives, and true negatives.Source: G. Performance Metrics · p. 7

H. Explainable AI integration

I. Simulation Setup

The evaluation describes a federated dashboard and reports 93% accuracy, alongside changes in precision, recall, and F1-score over training.

I. Simulation Setup

A. Web-based framework

  • The dashboard displays client training, updating, and idle states, counts gradient updates, and reports accuracy, precision, recall, and F1-score for each iteration.Its real-time accuracy-over-updates graph reports a score of 93%.Source: A. Web-based framework · p. 7

B. Model Performance Metrics

C. Explainable AI Insights

SHAP plots explain model predictions by showing how feature values affect them relative to a baseline. The conclusion describes decentralized training that keeps raw data local and uses SHAP to clarify influential features.

C. Explainable AI Insights

  • In Fig. 9, red/blue SHAP values indicate feature contributions that raise/lower predictions relative to the average-output baseline; larger magnitudes indicate stronger influence.The paper notes that features 2.15, 0.418, and 0.517 in Fig. 9b have more impact than 1.15 and -0.8.Source: C. Explainable AI Insights · p. 8, Figure 9 · p. 9

VI. CONCLUSION

  • Clients train locally and send model updates; Flask communication, multithreading, and client-weight aggregation update a central global model.Client weights are multidimensional arrays representing learned neural-network parameters.Source: §VI · p. 9
  • SHAP highlights influential features to help make federated decisions understandable and justifiable when centralized data collection is impractical or undesirable.The conclusion presents this as a way to build trust in sensitive domains.Source: §VI · p. 9
Loading 2312.13334v1…