AI for Fraud Detection in Financial Services
Financial fraud has evolved significantly with digital transformation, generating losses exceeding $41 billion annually across payment networks. Modern fraud schemes operate at scale, leveraging automated tools to process thousands of transactions per hour while adapting to detection patterns—making traditional rule-based systems increasingly ineffective at identifying sophisticated attack patterns.
The fundamental challenge lies in detecting fraudulent activities in real-time while maintaining acceptably low false positive rates across billions of daily transactions.
This page brings together solutions from recent research—including behavioral modeling systems that adapt to user contexts, distributed fraud detection architectures that preserve privacy, multi-tiered detection frameworks that identify group-level patterns, and reinforcement learning approaches for cash return fraud. These and other approaches focus on practical implementation in high-volume transaction environments while maintaining computational efficiency.
1. Machine Learning System with Categorical History Module and Decay Logic for Transaction Sequence Anomaly Detection
FEATURESPACE LIMITED, 2024
Machine learning system for processing transaction data that improves anomaly detection in transaction sequences. The system has a categorical history module with decay logic and update logic stages. It stores state data for transactions by category and entity, decaying older data based on time difference. New transaction input is updated and stored. The decayed and updated states are combined to output a scalar value representing anomaly likelihood in sequence. This allows detecting sequential patterns of abnormal transactions.
2. Automated Classification System for Anomalous User-Specified Values in Financial Transactions Using Feature Vector Analysis
Capital One Services, LLC, 2024
Automated detection of anomalous user-specified values in financial transactions to prevent fraud and errors. The method involves training a model to classify whether a user-specified value for a financial transaction is anomalous. When a transaction is verified, the user-specified value is compared to the verified value to generate a feature vector. This vector is fed into the trained model to classify as anomalous or non-anomalous. If classified as anomalous, the user is alerted and allowed to dispute the transaction to prevent execution. This allows catching errors and fraud in user-specified values without requiring a source of truth for the correct values.
3. Fraud Detection and Control System for Multi-Tiered Centralized Processing Using Clustering and Tier-Level Resource Locking
Alegeus Technologies, LLC, 2024
Detecting and controlling fraud in multi-tiered centralized processing systems like electronic benefits accounts. The system uses clustering, modeling, and locking to detect and prevent fraud at the group level. It clusters transactions by intermediary, generates models from historical data, detects fraud in a tier based on the models, and locks resources in that tier if fraud is found. This stops fraud at the tier level rather than just the transaction level to prevent masking in groups.
4. Transaction Monitoring System with Machine Learning-Based Bio-Behavioral Model Updates
SecureAuth Corporation, 2024
A system for detecting abnormal user transactions and identifying potential fraud using machine learning to continuously monitor user behavior and context. The system captures contextual and behavioral factors of users at a smart data hub to develop bio-behavioral models through machine learning. When a transaction request comes in, recent user data is sent to the hub to update the models. The updated models are then compared to the transaction factors to determine abnormality and risk. This allows identifying abnormal transactions based on deviations from a user's normal behavior and context.
5. Deep Reinforcement Learning Model with Dual-Network Architecture for Cash Return Fraud Recognition
ADVANCED NEW TECHNOLOGIES CO., LTD., 2024
Training a cash return fraud recognition model using deep reinforcement learning to accurately identify fraudulent cash returns in transactions. The model uses a two-network architecture with a primary network for recognition and a secondary network for training. The primary network takes transaction features as input to predict if it's a cash return. The secondary network takes the primary prediction along with transaction details to calculate a label value. This label value is used as the return for reinforcement learning to train the secondary network. The primary network's weights are periodically transferred to the secondary network for improved recognition. This allows the model to learn transaction features that indicate high-value cash returns.
6. Decentralized Fraud Detection Model with Group-Based Transaction Classification
Capital One Services, LLC, 2024
Distributing fraud detection for consumer transactions to groups of customers rather than using a centralized model. The method involves training a fraud detection model using multiple customer purchase histories to detect fraud and classify transactions. The model is then distributed to groups of customers, who can use it to detect fraud for their own transactions without sending data to a central server. This enables localized, decentralized fraud detection that is more scalable and private compared to a centralized model.
7. Neural Network-Based Fraud Detection System for Decentralized Electronic Payment Networks
RIPPLE LUXEMBOURG S.A., 2024
Detecting fraud in a decentralized electronic payment network using machine learning. The system estimates the route of a payment based on origin and destination, inputs it into a neural network to determine fraud probability, and takes action against high-risk payments and systems involved in fraud.
8. Fraud Risk Assessment System and Method for Digital Transactions with Third-Party Account Data Analysis
Plaid Inc., 2024
System and method for assessing fraud risk in digital financial transactions involving third-party accounts. It uses secured authentication and inspection to analyze account data for interaction assessment. The method involves accessing a user's external account, retrieving data, analyzing factors like identity, funding, transaction history, and assessing risk. This score is used to determine action like augmenting, flagging, or denying transactions based on fraud potential.
9. Fraudulent Merchant Detection Using Machine Learning with Individual and Transaction Scoring and Cohort Clustering
Stripe, Inc., 2024
Detecting fraudulent merchants in a commerce system that provides financial processing services for merchants and their agents. The fraud detection involves using machine learning models to analyze merchant activity and characteristics to identify fraudulent merchants. The models can include individual merchant scoring, transaction scoring, and cohort clustering and scoring. The individual merchant scoring involves training models to predict fraud based on merchant attributes and transaction history. The transaction scoring involves training models to predict fraud based on transaction attributes. The cohort clustering involves clustering merchants into groups based on shared attributes and scoring the clusters to predict fraud. Periodically analyzing merchant activity over time allows identifying cohorts associated with fraud even if individual merchants are not flagged.
10. Fraud Detection Engine Utilizing Multi-Model Analysis and Cohort Scoring in Commerce Systems
Stripe, Inc., 2024
Detecting fraudulent merchants using machine learning techniques in a commerce system. The system involves a fraud detection engine that analyzes merchant activity to identify potential fraud. The engine uses multiple fraud detection models that analyze different aspects of merchant behavior, like transaction patterns, agent usage, and decline rates. It also clusters merchants into cohorts based on similar attributes and scores those cohorts for fraud risk. This redundant check using cohort scoring helps catch merchants that may not be flagged by individual models. The fraud detection is performed periodically and asynchronously to analyze historical data and identify fraud trends.
11. Fraud Detection Model Training Using Authenticated User Action Data Acquisition
RAKUTEN GROUP, INC., 2024
Simplifying creation of fraud detection models in services by using authenticated user actions to train the models instead of manually labeled data. The system acquires authenticated user actions from services after authentication, and creates fraud detection models using those authenticated actions to learn what valid actions look like. This leverages known good user behavior to train the models instead of requiring manual labeling of training data.
12. Anomalous Activity Detection Method Integrating Transaction and Network Analysis with Combined Anomaly Scoring
DISCAL NV, 2024
A method for detecting anomalous activity, like fraud or money laundering, using a dynamic approach that combines transaction analysis and social network analysis. The method involves calculating anomaly scores for user transactions using unsupervised and supervised algorithms trained on transaction attributes. It also calculates network anomaly scores based on interconnected user profiles. The potential for anomalous activity is determined by combining the transaction and network scores. This provides dynamic detection from multiple angles rather than just transaction analysis alone.
13. Electronic Transaction System with AI-Driven Multimodal Data and Biometric Analysis
iWallet, Inc., 2024
Secure electronic financial transactions system that uses AI, biometrics, and multimodal data analysis to prevent fraud and improve user experience. The system collects multimodal data from visual, auditory, and tactile sensors during transactions. It uses AI modules like biometric authentication, transaction anomaly detection, geospatial analysis, behavioral analysis, and third-party data integration to analyze this data for fraud prevention. The system also communicates with users through modalities like vision, audio, touch, taste, smell, temperature, pain, and balance to address fraud concerns or request verification.
14. Distributed Register System for Immutable Transaction Data Storage and Machine Learning-Based Dispute Analysis
BANK OF AMERICA CORPORATION, 2024
Securely storing and analyzing transaction data using a distributed register and machine learning for dispute resolution. The system allows secure, immutable storage of resource transactions using a distributed register. Disputes can be initiated by users, analyzed using machine learning, and resolutions determined based on historical data. The distributed register prevents unauthorized data manipulation. Users can access verified data via multiple channels.
15. Electronic Transaction System with Fraudulent Phone Number Detection and Multi-Party Verification Mechanism
Visa International Service Association, 2024
Preventing fraud in electronic transactions by connecting merchants with known fraudulent phone numbers and rejecting transactions from similar numbers that have been flagged. The system also verifies transactions with separate responsible parties to combat spoofing where fraudsters change their displayed number.
16. Transaction Risk Assessment System Using Historical Data and Predictive Account Balance Modeling
SardineAI Corp., 2024
Reducing risk of transactions like ACH transfers by predicting account balances at settlement time using historical data and machine learning models. When a request is received to initiate a transaction, the system retrieves the account history and uses a trained model to project the balance at settlement. This allows assessing the risk of the transaction completing successfully without needing real-time account balance checks.
17. Fraud Detection System Utilizing Dual Machine Learning Models for Novelty and Similarity Analysis
Actimize LTD., 2024
Identifying fraud transactions in transactions classified as legit by a ML model using separate trained ML models for novelty detection. The method involves training separate ML models, one for fraud and one for legit transactions, using samples of historical labeled data. After deployment, transactions classified as legit by the main ML model are sent to the legit model to mark as similar or novel. Novel transactions are further sent to the fraud model. High novelty transactions are reported as potential unknown fraud and high similarity transactions are reported as potential missed fraud.
18. Graph Convolutional Neural Network System for Real-Time Anomaly Detection with Feedback-Driven Model Retraining
OPTUM, INC., 2024
Real-time anomaly detection using graph convolutional neural networks (GCNs) that can handle large-scale graph databases for applications like fraud detection. The method involves retraining the GCN model using feedback from manual anomaly confirmations to improve accuracy without offline redesign. When an anomaly is detected, a confirmation is performed to validate. Feedback is integrated into the training data to enhance the model. This allows real-time anomaly detection without requiring expensive redesigns or retraining when new graphs are added.
19. Autonomous System for Fraud Trend Detection and Mitigation Using Machine Learning-Based Rule Generation and Testing
JPMORGAN CHASE BANK, N.A., 2024
Autonomous fraud risk management system that quickly identifies and mitigates emerging fraud trends using machine learning and automation. The system performs feature engineering, rule recommendation, testing, and implementation in a closed loop process. It leverages machine learning techniques to develop fraud rules based on features extracted from data. The rules are tested in silent mode, approved, and upgraded to production. This automated and adaptive system allows efficient and timely creation of fraud rules to combat changing fraud trends.
20. Emotion Analysis-Based Fraud Detection System with Emotion Indexing and Variance Classification Models
Wells Fargo Bank, N.A., 2024
Detecting potentially fraudulent customer communications using emotion analysis. The technique involves analyzing the emotional content of customer communications to identify fraud. It involves two models: an emotion-based indexer to extract emotion factors from communications, and an emotion variance model to classify communications as fraud based on changes in emotional factors compared to historical ones. The emotion variance model determines if specific patterns of emotional factors indicate fraud. By analyzing emotional patterns, this technique aims to improve fraud detection accuracy compared to traditional methods.
Get Full Report
Access our comprehensive collection of patents related to this technology