Year-over-Year Developments in Financial Fraud Detection via Deep Learning: A Systematic Literature Review

1. Introduction

• Fraud spans credit cards, insurance, and money laundering, causing financial losses and eroding trust in financial systems.Global estimates suggest organizations lose 5% of annual revenues to fraud.Source: §1 · p. 1
• Manual reviews and rule-based systems are increasingly inadequate against sophisticated schemes and the surge in digital transactions.The passage reports more than 2.7 billion U.S. digital transactions in 2023.Source: §1 · p. 1
• The study explores recent machine-learning techniques for fraud detection, focusing on applications, effectiveness, compliance, and persistent data-quality, interpretability, and ethical challenges.The introduction names CNNs, LSTMs, and NLP as advanced techniques for real-time anomaly detection and adaptation to evolving threats.Source: §1 · pp. 1–2

2.1. Study Design

• The Kitchenham framework structures study selection, data extraction, and synthesis to examine gaps, trends, and challenges in this interdisciplinary field.The authors describe the approach as supporting transparency, replicability, and unbiased results.Source: §2.1 · p. 2

2.2. Research Questions

• The research questions address fraud-sector trends, preprocessing and automation, model advances, evaluation benchmarks, and privacy and regulatory change.They are presented as aligned with the study’s objectives for researchers and practitioners.Source: §2.2 · p. 2

2.3. Search Criteria

• The search covered six databases and combined Boolean keywords for methods, financial sectors, and fraud.The databases were PubMed, SSRN, IEEE Xplore, ACM Digital Library, ScienceDirect, and Scopus.Source: §2.3 · pp. 2–3
• The search excluded review and survey papers and restricted subject areas to fields at the intersection of technology and finance.The targeted areas included computer science, business, economics and finance, decision sciences, and engineering.Source: §2.3 · p. 3

2.4.1. Inclusion Criteria

• Eligibility focused on peer-reviewed studies published from 2019–2024 that use deep learning for financial fraud detection.The inclusion criteria name CNNs, RNNs, LSTMs, and transformers as examples.Source: §2.4 · p. 3, §2.4.1 · p. 3

2.4.2. Exclusion Criteria

• The review excluded non-English articles, studies outside financial services, and studies without deep-learning techniques.Source: §2.4.2 · p. 3

2.5. Data Extraction and Analysis

• A structured extraction form standardized data collection and synthesis across studies to support clear comparisons and minimize subjectivity.Source: §2.5 · p. 3
• Python libraries supported data manipulation, visualization, and statistical tasks, while VOSviewer mapped keyword co-occurrence networks.The passage names Pandas, Matplotlib, and Scikit-learn for the Python analysis.Source: §2.5 · p. 3

3. Results

• 2,858 initial search results were narrowed to 427 papers for further evaluation and then 57 papers meeting the review’s objectives and standards.The screening process is shown in Figure 1.Source: §3 · p. 4, Figure 1 · p. 4

3.1. Research Question 1

• Research-paper counts increased steadily from 2019 to 2021, rose significantly from 2022, and climbed steeply from 2023 to 2024.The passage suggests technology advances, fraud concerns, and policy regulation as possible drivers, not established causes.Source: §3.1 · p. 4, Figure 2 · p. 5
• Credit card and banking are identified as the two main sectors contributing to the publication increase.Source: §3.1 · p. 5
• Credit-card research is prominent, and the review suggests public datasets such as the European Credit Card Transactions dataset may have contributed to that concentration.The passage describes the dataset as clean and standardized for model development and evaluation.Source: §3.1 · p. 5, Figure 3 · p. 6
• Banking and insurance research addresses digital payments, automobile claims, and health-insurance claims, while tax, mortgage or loan, and money-laundering studies are less numerous.The review suggests limited domain-specific data access or complex fraud patterns could contribute to the lower publication counts.Source: §3.1 · p. 5
• Crypto/blockchain and payment-system studies indicate growing research attention to digital-currency fraud prevention.The review also reports 2023 losses of $1.86 billion for bank transfers and payments and $1.41 billion for cryptocurrency.Source: §3.1 · p. 5

3.2. Research Question 2

• 48 of 57 reviewed papers reported dataset imbalance, which can make traditional models favor the majority class and underperform on fraud cases.Source: §3.2 · p. 6

3.2.1. Preprocessing Techniques

• SMOTE increases minority-class diversity and has also been reported effective for imbalanced datasets with missing values [30].Source: §3.2.1 · p. 6
• Timestamp-based stratified sampling represents more recent fraud cases while retaining randomness, addressing non-stationary fraud characteristics.Source: §3.2.1 · p. 6
• GANs and VAEs generate realistic synthetic data that preserves the original distribution for extreme imbalance, while kNN imputation addresses missing values.Source: §3.2.1 · p. 6
• Scaling and normalization reduce bias, while ADASYN targets harder-to-classify cases and cluster-based oversampling tailors synthetic data to domain contexts.Source: §3.2.1 · p. 6

3.2.2. Automation Techniques

• Automation reduces manual intervention and costs; subsampling lowers computational costs while preserving feature correlations, and VFDT with blockchain supports real-time processing and secure, scalable updates.Source: §3.2.2 · p. 7
• Blockchain supports automated detection through smart contracts, while incremental SGD, automated tuning, and hybrid resampling are reported to enhance detection accuracy.Source: §3.2.2 · p. 7
• Knowledge distillation supports efficient real-time detection, while stacking ensembles improve robustness and the Random Forest Quantile Classifier optimizes sensitivity and specificity for imbalanced data.Fraud detection pipelines combine preprocessing and model training into cohesive systems.Source: §3.2.2 · p. 7

3.3.1. Deep Learning Models

• CNNs capture features in transaction data, while GRU and LSTM mitigate vanishing and exploding gradients in RNNs; Transformers process sequence inputs in parallel using self-attention.CNN-derived features also enhance fraud detection when combined with traditional models.Source: §3.3.1 · pp. 7–8
• NLP methods use textual financial, regulatory, and claims data; GANs and VAEs offer different sample-generation and representation-learning strengths, while GCNs outperform LR, RF, and GBMs in one study.The review notes that GANs are harder to train and less interpretable, whereas VAEs have a structured latent space but generate lower-quality samples; combining them may address imbalanced-data limitations.Source: §3.3.1 · p. 8

3.3.2. Machine Learning Models

• Traditional models serve as benchmarks: Random Forests resist overfitting, Gradient Boosting handles imbalanced data, and interpretable Decision Trees remain prone to overfitting.LR offers simplicity and interpretability, while KNN detects anomalies by proximity but lacks scalability for large datasets.Source: §3.3.2 · p. 8

3.3.3. Hybrid Models

• RDQN combines a DNN with reinforcement learning and Rough Set feature reduction, achieving faster processing and higher accuracy while outperforming traditional and hybrid comparators.The cited comparisons include SVM, ANN, DT, IFDTC4.5, SAE-GAN, and CNN-SVM-KNN.Source: §3.3.3 · p. 9
• Transformer-LOF-Random Forest surpasses XGBoost, LightGBM, and LSTM; RXT-J outperforms BERT, ANN, and logistic regression; CatBoost-DNN exceeds random forests and LSTM-based AdaBoost.The models combine Transformer feature extraction with local anomaly detection and ensemble learning, ResNeXt with GRU, and CatBoost with DNN, respectively.Source: §3.3.3 · p. 9
• Autoencoder-LSTM combines dimensionality reduction and noise removal with temporal modeling, and outperforms traditional machine-learning methods and standalone LSTM.The autoencoder retains key features; the LSTM models temporal dependencies and classifies transactions as fraudulent or legitimate.Source: §3.3.3 · p. 9

3.3.4. Trends in Model Usage

• LSTM, MLP, CNN, and RNN are widely used; LSTM use rose sharply from 2022 to 2024, and credit-card and banking applications commonly use MLP, LSTM, and CNN.Tax and money-laundering sectors show minimal deep-learning application; GAN use is emerging for synthetic data addressing imbalance.Source: §3.3.4 · pp. 9, 10, 12, Figure 8 · p. 12
• Traditional machine-learning models are frequently connected to deep-learning keywords as baselines or hybrids; MLP is the most relevant deep-learning keyword, while GAN is used to generate synthetic data for imbalanced datasets.MLP has 27 links and 103 link strength; Random Forest has 32 links and 234 link strength, and Logistic Regression has 31 links and 191 link strength.Source: §3.3.4 · pp. 11–12

3.4. Research Question 4

• Accuracy can mislead on imbalanced data; Precision may reduce Recall, high Recall can increase false-positive rates and costs, and F1 balances Precision and Recall without measuring economic impact.AUC-PR is described as more relevant than AUC-ROC for imbalanced data because it focuses on fraud-detection effectiveness.Source: §3.4 · p. 12
• Economic metrics capture operational and financial impacts, with Precision prioritized in cryptocurrency fraud and Recall in tax fraud.The stated aims are minimizing compliance costs and preventing revenue losses, respectively.Source: §3.4 · p. 13
• Effective evaluation increasingly tailors metrics to imbalanced data, operational costs, and sector-specific needs alongside foundational measures such as Accuracy and F1 Score.Source: §3.4 · p. 13

• 36 reviewed papers used publicly available datasets, 24 used nonpublic datasets, and 19 addressed privacy.Source: §3.5 · p. 13
• PCA reduces data dimensionality and anonymizes transactions for secure inter-institutional sharing, but may reduce interpretability and insight into variable relationships.Source: §3.5 · pp. 13–15
• Blockchain uses encrypted, decentralized ledgers and pseudonymous addresses to protect identities and transaction data.Source: §3.5 · p. 15
• GDPR’s right to explanation requires interpretable fraud-detection decisions [34], while CCPA grants access, deletion, and opt-out rights that constrain data use.Source: §3.5 · p. 15

4.1. Credit Card

• Credit-card fraud research benefits from publicly available large-scale datasets; deep learning models can process high transaction volumes quickly for real-time or near-real-time decisions.Source: §4.1 · pp. 15–16

4.2. Insurance

• Insurance data exchanges raise breach and inaccuracy risks under HIPAA and GDPR safeguards; federated learning supports collaborative training without exposing sensitive data.Source: §4.2 · p. 16

4.3. Blockchain

• Combining blockchain with XGBoost and VFDT increased detection accuracy and reduced error rates by 7% compared to traditional methods.Source: §4.3 · p. 16
• Permissioned blockchains support secure data sharing and transparency, while smart contracts streamline claims processing and reduce errors and processing time.Source: §4.3 · p. 16

4.4. Banking & Payment

• Banking fraud spans online banking and diverse payment channels; challenges include real-time detection, transnational networks, and integrating sources such as KYC profiles.Source: §4.4 · p. 16

• The 2019–2024 window excludes earlier foundational work, and public-literature selection omits proprietary models and industry practices.Source: §5 · p. 17
• English-only coverage may miss regional advances, while inconsistent data processing and evaluation frameworks make cross-study comparisons challenging.Source: §5 · p. 17
• Future work should standardize preprocessing, including imbalance handling, and domain-specific benchmarks that include economic and operational metrics.Source: §5 · p. 17
• The proposed lifecycle agenda also calls for scalable deployment with real-time detection and privacy compliance, plus adaptive monitoring and retraining as fraud and regulations change.Source: §5 · p. 17

• CNNs, LSTMs, transformers, and ensemble techniques have enhanced detection of complex fraud patterns across diverse financial sectors.Source: §6 · p. 17
• Blockchain, PCA, and GDPR and CCPA compliance align fraud-detection advances with ethical and legal standards while protecting sensitive data.Source: §6 · p. 17
• Parameter tuning and subsampling are reported to accelerate time-to-detection, making real-time fraud mitigation feasible in high-stakes environments.Source: §6 · p. 18
• Future research should improve interpretability, address emerging fraud schemes, and expand cross-industry collaboration through federated learning and secure data sharing.Source: §6 · p. 18