Financial Fraud Prevention using AI

This review examines the impact of using Federated Learning (FL) on No-Code AI tools and how it could change data security in BFSI industry. Despite keeping local each machine, FL allows different organizations to train models together. It gives an overview what components model contains, its functions accurate is at predicting things by comparing other machine learning that were used as references. Our work highlights successful cases uses innovative develop approach can boost prevention, analysis monitoring fraud, risks compliance BFSI. The challenges included paper are heterogeneous data, threats challenge scaling, along with future research ideas. These findings matter most specialists working AI, experts finance, who need more privacy their solutions.

Source

This study aims to compare the performance of two popular machine learning algorithms, Random Forest and K-Nearest Neighbor (KNN), in predicting creditworthiness online lending systems. The research uses publicly available Loan Approval Prediction Dataset from Kaggle, which contains borrower profiles such as employment status, number dependents, annual income, loan amount, term, credit score. Data preprocessing included cleaning, handling missing values, outlier removal, transformation through normalization encoding. dataset was divided into 80% training data 20% testing data. configured with 100 decision trees unlimited depth, while KNN used an optimal k value 5 determined by grid search. Model evaluated using accuracy, precision, recall, F1-score. results showed that outperformed consistently higher values (97%) across all metrics, demonstrating strong stability superior pattern recognition capabilities. KNN, accuracy 89%, still good can be considered a lightweight alternative for simpler applications.

Source

This study aims to develop a mechanism for cyclic-dynamic screening of the activities primary financial monitoring entities (PFMEs) under wartime conditions. The proposed identifies atypical behavioural patterns, enhances responsiveness rapidly evolving risks, and improves quality offence detection. constructed additive time series models two distinct periods: pre-war (20112019) (20212024). Smoothing techniques, decomposition, seasonal component estimation, analytical trend modelling were applied. In first period, model demonstrated high accuracy (coefficient determination R = 90,87%), while in second remained acceptable (R 65,80%), reflecting increased volatility environment. results highlight importance analyzing fluctuations trends detect suspicious transactions on improve effectiveness procedures during wartime. identifying anomalous operations, enables adaptive risk response, strengthens control flows. practical implementation this methodology contributes ensuring national stability security, which is critically important maintaining integrity state's system

Source

Phishing and digital arrest attacks pose significant threats to e-commerce security, necessitating advanced AI-driven detection mechanisms. In this context, study proposed a hybrid deep learning model integrating BERT for feature extraction, Atrous Spatial Pyramid Pooling (ASPP) multi-scale refinement, CNN-based classification network. The encoder captures semantic contextual features, while ASPP enhances spatial hierarchies using dilated convolutions. A CNN classifier with batch normalization ReLU activation ensures robust classification. Extensive experiments on Kaggle dataset demonstrate superior performance, achieving 97.81% accuracy, 98% precision, an AUC of 0.9972, outperforming GRU, LSTM, RNN, Transformer models.

Source

Abstract: This study presents an automated framework for threat intelligence gathering using Natural Language Processing (NLP) on dark web data. The growing sophistication of cyberattacks necessitates real-time detection emerging threats. Traditional manual analysis forums is time-consuming and insufficient. research proposes a hybrid NLP pipeline that integrates named entity recognition, sentiment analysis, topic modeling to extract actionable indicators from darknet discussions. A dataset comprising over 100,000 posts was analyzed, yielding high accuracy in identifying cyber actors, malware variants, planned attacks. proposed model outperformed baseline models by 17% F1 score. findings highlight the utility systems reducing human workload accelerating defense responses. Keywords Threat Intelligence, Processing, Dark Web, Cybersecurity, Automation, Topic Modeling

Source

<title>Abstract</title> <italic>With the rapid growth of financial data, extracting accurate and contextually relevant information remains a challenge. Existing question-answering (QA) models struggle with domain-specific terminology, long-document processing, answer consistency. To address these issues, this paper proposes Intelligent-Aware Transformer (IAT), QA system based on GLM4-9B-Chat, integrating multi-level aggregation framework. The employs Financial-Specific Attention Mechanism (FSAM) to enhance focus key terms, Dynamic Context Embedding Layer (DCEL) improve Hierarchical Answer Aggregator (HAA) ensure response coherence. Additionally, Knowledge-Augmented Textual Entailment (KATE) strengthens models generalization by inferring implicit knowledge. Experimental results demonstrate that IAT surpasses existing in tasks, exhibiting superior adaptability long-text comprehension reasoning. Future work will explore computational optimizations, advanced knowledge integration, broader applications.</italic>

Source

Against the backdrop of rapid advancements in fintech and continuous expansion credit markets, traditional risk assessment methods have revealed significant limitations. Machine learning offer new opportunities for management. This study focuses on applying machine to We utilize a dataset from Kaggle (Default Credit Card Clients Dataset) analyze compare performance logistic regression, decision tree, random forest models across multiple dimensions, including accuracy, recall rate, interpretability. The results demonstrate that tree model exhibits comprehensive default prediction. Future research could incorporate diverse data types, develop visualization tools, establish real-time monitoring dynamic updating systems, extend applications industries enhance accuracy foresight assessment, thereby promoting widespread adoption financial holds theoretical significance offers practical technical solutions real-world lending operations.

Source

Extensible fraud detection system for service providers that allows users to create custom fraud detection rules and immediately implement them, reducing fraud without blocking legitimate transactions. The system groups rule features based on data sources to reduce latency, vets new rules on historical data to test effectiveness, and warns about overly aggressive rules. It also caches frequently accessed features to further reduce latency. This allows users to easily create and deploy custom fraud detection rules without negatively impacting transaction processing time.

Source

Reducing false positives in financial transaction fraud detection using machine learning. The method involves training a false positive reduction model (FPRM) recursively using labeled transaction data from an initial fraud scoring model. The labeled data includes transactions marked as high risk that were later determined to be either false positives or true positives. The recursive training allows refining the FPRM to better distinguish true fraud from false positives. The FPRM is then used in production to further reduce false positives in fraud detection.

Source

This paper is aimed at developing a computational intelligence model for real-time detection and prevention of credit card fraudulent transactions within digital cyber forensic investigations. Decision Trees, Support Vector Machines Artificial Neural Networks were employed in the design system to ensure reliable efficient fraud detection. In order eliminate noise enhance accuracy analysis, actual transaction data entered set was used. The trained through supervised learning technique identify patterns real time. To verify effectiveness developed system, post-hoc comparisons done regarding models terms accuracy, precision, recall, f1 score. calculation revealed that provide best as it reached 98% precision correct activity identification. research has helped reduce rise ecosystem by employing contemporary approaches. It also assists investigators mitigate financial damage enhances security measures institutions.

Source

Machine learning model framework for accurately classifying data like transactions using multiple graph analysis techniques to overcome limitations of single-dimensional models. The framework uses multiple models that analyze actual attributes, fuzzy attributes, and community aspects of transactions in a graphical manner. It generates a graph representing connections between transactions and fuzzy attributes. This allows capturing relationships and patterns among transactions that would be missed by single-dimensional models. By analyzing multiple dimensions, the framework classifies transactions more accurately compared to conventional models that analyze just one or two dimensions.

Source

Generating targeted transaction processing rules for fraud prevention using machine learning without requiring expert guidance. The system identifies predictive attributes from historical transactions that indicate fraud, allows a user to select one, and sets a threshold value for that attribute. This generates a customized rule to block future transactions with that attribute value. The system processes transactions using the customized rule to detect and prevent fraud scenarios specific to the user.

Source

Automated generation of suspicious activity reports (SARs) using large language models (LLMs) trained on transaction data and bank databases. The system extracts transaction features, predicts fraud risk, and generates an LLM prompt based on the transaction. It then generates a SAR report using the LLM to format and summarize the transaction data. This provides consistent and complete SARs compared to manual preparation.

Source

Unified Payments Interface (UPI) has revolutionized digital transactions in India offering convenience and real-time processing. However, this rapid adoption also led to a surge fraudulent activities, challenging the efficacy of traditional rule-based fraud detection methods. These conventional systems often struggle adapt evolving patterns, necessitating more robust adaptive solutions. In response these challenges, researchers have explored machine learning techniques detect activities within UPI transactions. While existing strategies shown promise, they are frequently validated on limited or synthetic datasets, which may not fully capture complexities real- world scenarios. To address limitations, comprehensive evaluation prevalent classifiersincluding Logistic Regression, K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Naive Bayes, Decision Trees, Random Forests, Convolutional Neural Networks (CNNs)was conducted. Building upon analysis, we propose (CNNs) framework, designed enhance capabilities

Source

Detecting suspicious transaction patterns in financial networks using graph neural networks trained on synthetic patterns. The method involves generating a transaction graph for a reference network with synthetic suspect patterns, extracting subgraphs, training a GNN on those subgraphs, and applying it to new networks to find potentially fraudulent patterns. The subgraphs are dynamically defined based on node metadata to include correlated activity.

Source

Detecting and preventing fraud in financial accounts by analyzing user behavior and patterns to identify potential fraudsters. The system analyzes factors like phone numbers, call origins, voice profiles, and question-answer sequences to determine likelihood of fraud. If fraud is suspected, alerts are sent to users and restrictions are placed on account access. The system uses AI/machine learning to train models on known fraud methods and behavior patterns.

Source

1.ABSTRACT Frauds in online transactions is increasing like crazy these days, and old methods just ain't cutting it anymore. Banks businesses losing tons of money cause the fraudsters always coming up with new tricks. But guess what? AI might be hero we need - its getting really good at spotting fakes before they happen.. This paper gonna look how machine learning stuff helps catch frauds better than humans can. We talk about systems learns from past data, finds weird spending patterns that don't make sense, even predicts type scams spread. not perfect sometimes blocks normal people's purchases (so annoying!) or misses clever frauds, Also there Is big questions privacy since needs so much personal data's to work good. Looking real cases, see already helping lots companies reduce by 40-50%. technologys still got ways go can fully trust it. Main point making payments safer before, but gotta fix mistakes smarter.

Source

Real-time transaction approval process that improves speed and stability of approving transactions by using named entity restrictions based on machine learning models. The process involves identifying entities involved in transactions from past data, creating rules specifying authorized entities, and iteratively training machine learning models to accurately identify entities in real-time. The identified entities are compared against the rules to approve transactions. This allows faster approval compared to manual review while still enforcing entity restrictions.

Source

Generating a virtual asset wallet address blacklist using graph attention networks (GAT) to proactively detect high-risk virtual asset wallets that are prone to illegality. The blacklist is generated by training a GAT model using full node indices, common transaction data, and an existing blacklist. The model calculates scores for wallet addresses based on their connections and transactions. Addresses with high scores are added to the GAT blacklist.

Source

Sharing transaction-related data between entities of a payment network using a common transaction ID. The method involves generating a confidence score indicating the likelihood that two transactions are the same using a machine learning model. If the confidence score exceeds a threshold, a common transaction ID is generated based on the attributes of one transaction. This common ID is then used to request additional transaction data from the other entity referencing the common ID.

Source