Fraud Detection in the Airline Industry: Tackling Financial Losses through Transfer Learning and Federated Learning

Fraud is a significant challenge in the airline industry, leading to billions of dollars in financial losses each year. According to the International Air Transport Association (IATA), in 2021 alone airlines faced an estimated €6.5 billion in losses due to fraudulent activities, which accounted for around 1.5% of the industry’s total revenue. This encompasses a variety of fraud types, including stolen credit card transactions, chargeback abuse, loyalty program fraud, and ticket scams, which not only reduce revenue but also undermine customer trust and raise compliance costs [1]. Additionally, a Deloitte report [2] indicated that airlines face around €440 million ($600 million) in annual losses due to fraudulent activities, which can range from fraudulent ticket sales to employee misconduct. Fraud not only leads to direct financial losses but also impacts customer trust, operational efficiency and transaction costs due to the need for heightened fraud prevention systems. As airlines increasingly rely on digital transactions and online booking systems, fraudsters are becoming more sophisticated in exploiting system vulnerabilities. Traditional rule-based fraud detection systems are insufficient for handling the complexity and volume of modern fraud patterns.

Artificial Intelligence (AI) provides an effective solution for automating fraud detection and identifying suspicious behaviors in the airline industry [3,4]. By leveraging machine learning models, AI can detect anomalies and recognize fraudulent patterns across various types of fraud. For instance, AI excels at identifying stolen credit card transactions, where unauthorized individuals use someone else’s credit card to book flights or services without consent. It also effectively addresses chargeback abuse, where customers fraudulently claim refunds for legitimate bookings by disputing the transactions. Additionally, AI can help uncover loyalty program fraud, where unauthorized access or manipulation of frequent flyer miles and rewards occurs, often leading to significant financial and reputational losses. AI models also play a crucial role in detecting ticket scams, where fraudsters trick customers into purchasing fake or non-existent flight tickets, causing financial harm to both the airline and its customers. Through advanced algorithms and continuous learning, AI enhances the speed and accuracy of fraud detection processes, providing airlines with more robust tools to prevent and mitigate fraud while maintaining customer trust and operational efficiency.

While AI can address these challenges, traditional models are often limited by the data they have access to and the learning capacity within a single environment [5,6]. This is where transfer learning and federated learning bring unique value. Transfer learning allows AI models to leverage knowledge gained from one domain and apply it to another, enabling more accurate fraud detection across different airline systems, even when data is scarce or isolated [7-9]. Federated learning takes this a step further by allowing multiple airlines or organizations to collaboratively train AI models without sharing sensitive data, ensuring privacy and security while still benefiting from the collective insights of vast and diverse data sources [7-11].

The combination of transfer learning and federated learning is particularly novel. Together, they allow AI models to continuously learn and adapt not only from their own environment but also from other environments and industries without compromising data privacy [7-11]. This dual approach enhances the robustness of fraud detection models by applying knowledge transfer across domains while benefiting from the shared insights of global industry players. This synergy creates a more adaptive, secure, and scalable solution for combating fraud in the airline industry, enabling models to stay ahead of evolving fraud tactics with greater accuracy and resilience.

In this research, we developed a fraud detection system that integrates transfer learning and federated learning. The system combines TabNet, a deep learning model optimized for tabular data, and BERT (Bidirectional Encoder Representations from Transformers), a cutting-edge Natural Language Processing (NLP) model. TabNet excels at handling structured data such as transactions and behavioral patterns, while BERT, developed by Google, analyzes the context of words bidirectionally to capture deeper semantic meaning, making it highly effective for textrelated tasks like classification and inference. By combining these two models, our system effectively detects multiple types of fraud, including stolen credit card transactions, chargeback abuse, loyalty program fraud, and ticket scams. The synergy between TabNet’s handling of transactional data and BERT’s natural language processing capabilities allows for a comprehensive approach to fraud detection across different data types, improving overall accuracy and efficiency.

The results have demonstrated that the system exhibited strong performance across various fraud categories in the airline industry. Initially, the model showed strong results, achieving 95% accuracy in detecting stolen credit card transactions, 98% for chargeback abuse, 93% for loyalty program fraud, and 96% for ticket scams. Precision and recall scores were similarly high, with stolen credit card fraud reaching 92% precision and 93% recall, while ticket scams achieved 94% for both metrics. After implementing federated learning across two airline nodes, performance further improved. Accuracy increased to 96% for stolen credit card transactions, 99% for chargeback abuse, 94% for loyalty program fraud, and 97% for ticket scams. Federated learning also enhanced precision and recall, with stolen credit card fraud reaching 93% precision and 94% recall, and ticket scams improving to 95.3% precision and 97% recall. These findings underscore the effectiveness of integrating transfer learning with federated learning to develop a scalable, accurate, and privacy-preserving fraud detection solution tailored for the airline industry.

The remainder of the paper is structured as follows: Section 2 presents a literature review, Section 3 outlines the methodology used for prototype development, Section 4 details the datasets utilized, and Section 5 reports the results. Section 6 offers a discussion, and Section 7 concludes the paper.

We conducted a literature review across various academic sources to investigate the use of federated and transfer learning for fraud detection in the aviation industry. The review revealed that information on this specific topic is notably scarce. While some studies have explored related areas, such as machine learningbased fraud detection systems and privacy-preserving models in aviation, the direct application of federated and transfer learning to aviation fraud remains largely unexplored. This underscores a significant research gap and presents an opportunity for future studies to apply these advanced techniques in the aviation fraud detection field.

Literature review

As part of MT Aras & MA Güvensan [3], the authors developed a machine learning-based fraud detection system for the aviation industry. They employed several techniques, including under sampling to balance fraud and legitimate records, feature selection to reduce unnecessary data, and cost-sensitive metrics to evaluate the system’s performance. The study focused on detecting zero-day fraud attacks and experimented with various models, including Random Forest and SVM. The most prominent frauds detected were stolen credit card transactions, and the system showed improved recall by 24% compared to rule-based systems. Experimental results highlighted the significance of balancing detection rates with cost-sensitive performance.

Authors in Prathim C et al. [4] proposed a zero-trust model enhanced by blockchain-based identity storage and advanced algorithms to strengthen airline communication network security. They applied blockchain to decentralize and secure identity management, reducing the risk of identity theft. Machine learning and anomaly detection algorithms were used to identify suspicious behavior in real-time. This model was designed to combat data breaches, unauthorized access, and cyberattacks. Experimental results showed the model effectively minimized security risks, while advanced encryption techniques ensured the integrity of data exchanges in the airline industry.

As part of Zhou X et al. [5], the authors proposed a supervised privacy preservation transaction system for aviation business using a consortium blockchain. The system utilized identity-based homomorphic encryption, where both the airline and a supervisor could decrypt transaction amounts independently without interaction. This model aimed to enhance transaction security by preventing privacy leaks and addressing fraud risks, such as unauthorized access to financial records. The experimental results demonstrated encryption and decryption efficiency, with times of 15ms and 15.45ms for the dual recipients. The system provided strong privacy protection while enabling supervision.

Authors Zhuang Z et al. [6] developed a symbolic classifier using Genetic Programming (GP) to detect turbulence anomalies in aviation. They utilized Quick Access Recorder (QAR) data for this purpose. The classifier bypassed the need for traditional Eddy Dissipation Rate (EDR) calculations. Experimental results demonstrated the classifier’s accuracy, achieving a 98.08% accuracy rate in detecting turbulence, particularly effective in identifying severe turbulence conditions. The proposed method outperformed other machine learning algorithms like Random Forest and SVM in robustness and general applicability across different flight routes.

As part of Wang M et.al [12], the authors proposed an AGIfedavg algorithm to improve federated learning for UAV-based power line inspections. This algorithm enhanced the Fedavg algorithm by addressing the non-IID data problem, accounting for data disparities across nodes. The authors demonstrated the algorithm’s effectiveness using the MNIST dataset. AGI-fedavg achieved superior accuracy (0.9725) compared to Fedavg (0.97) and Fedprox (0.9571), particularly excelling in communication efficiency and model performance. The method successfully protected data privacy and enabled collaborative learning without sharing raw data.

We applied a combination of transfer learning and federated learning in the context of fraud detection for various types of airline fraud, including stolen credit card transactions, chargeback abuse, loyalty program fraud, and ticket scams. By utilizing pretrained models, we enhanced them with airline-specific data, while maintaining privacy through federated learning, which enables collaborative model training without centralizing sensitive data.

Transfer learning with pre-trained models

For this research, we used two distinct models, TabNet and BERT, which were integrated to handle structured and unstructured data effectively. TabNet is specifically designed for tabular data, which makes it highly suitable for processing structured transactional and behavioral data, such as payment logs and customer profiles. One of Tab Net’s key advantages is its ability to learn important features for fraud detection without the need for manual feature engineering. It supports anomaly detection (e.g., for stolen credit card transactions and ticket scams), pattern recognition (e.g., for chargeback abuse), and behavioral analysis (e.g., for loyalty program fraud). BERT, specifically the BERT-Base, uncased model, was pre-trained on BookCorpus giving it a broad understanding of general language semantics. We fine-tuned this pre-trained model on unstructured text data relevant to fraud detection, such as transaction descriptions, customer communications, and chargeback disputes. BERT is especially useful in analyzing natural language data for understanding complex linguistic patterns and identifying potentially fraudulent text-based behavior.

Federated learning for collaborative model training

After fine-tuning the TabNet and BERT models using generic fraud detection data, we implemented federated learning to extend the system’s capabilities by enabling collaboration between multiple airlines while ensuring data privacy. For this prototype, we created two nodes, each representing a different airline:
A. Each airline locally trained fraud detection models using its own data, including transactional logs, customer complaints, and other relevant information.
B. Instead of sharing raw data, only the model updates (weights) were exchanged between the nodes, refining the system’s overall fraud detection performance.

This approach allowed the system to learn from diverse data sources while maintaining strict privacy protections, ensuring that no sensitive customer information was shared between the airlines. By utilizing federated learning, we were able to enhance the system’s ability to detect fraud across multiple airlines without compromising data security.

Example of the combined approach

Our combined approach leverages both TabNet and BERT models within a federated learning framework. Here’s how the process works:
a) TabNet was applied for detecting transaction-based fraud such as stolen credit card use and ticket scams by analyzing structured behavioral data.
b) BERT was utilized for analyzing textual information, such as emails or chargeback disputes, to detect patterns indicative of fraud.
c) The models were fine-tuned using data from a single airline.
d) Federated learning was then applied to enhance the model by collaborating across multiple airlines, improving robustness and accuracy over time without compromising data privacy.

By combining transfer learning and federated learning, we created a flexible, scalable fraud detection system that improves continuously as it learns from diverse sources, making it highly effective across all fraud types in the airline industry. This system demonstrated robust performance in detecting stolen credit card transactions, chargeback abuse, loyalty program fraud, and ticket scams, achieving high accuracy while ensuring data privacy and compliance.

For developing the fraud detection prototype we used several dataset that targeted different types of fraud. Each dataset was carefully chosen to align with the specific fraud detection tasks, allowing us to apply a combination of transfer learning and federated learning effectively while ensuring data privacy.

Datasets for fraud types

The following datasets were used for the prototype:
Stolen Credit Card Fraud: The Credit Card Fraud Detection Dataset from Kaggle [13] was used to detect stolen credit card fraud. This dataset contains transactional data with labels indicating whether a transaction is fraudulent, making it ideal for identifying unusual spending patterns.
Chargeback Abuse: For chargeback abuse detection, we used the world bank open dataset [14]. This dataset contains textual data, including customer complaints, which were processed using Natural Language Processing (NLP) models like BERT to detect fraud-related patterns in chargeback disputes.

To simulate loyalty program fraud, we utilized customer churn Dataset from Kaggle [15] and then transformed to Loyalty Rewards Fraud Detection Dataset. The transformed dataset includes synthetic transactional, and loyalty point data, helping to detect fraudulent point accumulation and redemption activities.

For detecting ticket scams, we sourced data from the Airline Booking Datasets from OpenML [16] and Airline Data from Data. gov [17]. These datasets provided valuable transactional and behavioral data, enabling the system to recognize unusual booking patterns.

Data splitting for transfer learning

For transfer learning, we began by selecting pre-trained models: TabNet for tabular transactional data and BERT-Base, uncased for textual data. These models were fine-tuned on airlinespecific datasets, optimizing them for detecting industry-specific fraud patterns. The data was split into:
A. Training set: 70-80% of the dataset was used for training the model to learn airline-specific fraud patterns.
B. Validation set: 10-15% of the dataset was used for finetuning hyperparameters and validating the model during training.
C. Test set: 10-15% of the dataset was held out to evaluate the model’s performance after training, ensuring robust results.

This split allowed the models to adapt their pre-trained knowledge (from large, generic datasets) to the specific fraud detection needs of the airline industry.

Federated learning setup

We implemented federated learning to enhance the system across multiple airlines while maintaining data privacy. For the prototype, we created two nodes representing two airlines. Each airline used its own local dataset to train the models:
A. Node 1 (Airline A): Trained locally on a subset of the dataset (e.g., European transactions or loyalty program data).
B. Node 2 (Airline B): Trained locally on another subset of the dataset (e.g., American transactions or chargeback disputes).

The federated learning process involved:
a) Each node (airline) fine-tuned its models using its own data without sharing any raw data.
b) Only the model updates (weights) were shared between nodes and aggregated using federated averaging. This allowed the global model to learn from the experiences of multiple airlines, enhancing its ability to detect diverse fraud patterns across regions and fraud types, while preserving the privacy of each airline’s sensitive data.

Example: Chargeback abuse detection

For detecting chargeback abuse, we used the World Bank Open DataSet [14] containing textual data that was analyzed using the BERT model. The following outlines the methodology we employed:
Transfer learning: We fine-tuned the pre-trained BERT model using customer complaint data, identifying patterns related to fraudulent chargeback disputes. The dataset was split into 70% training, 15% validation, and 15% test sets to fine-tune the model for this specific task.
Federated learning: After local fine-tuning at each airline (e.g., Airline A and Airline B), we applied federated learning:
A. Node 1 (Airline A): Used 25% of the dataset, focusing on transactions from a specific region (e.g., Europe).
B. Node 2 (Airline B): Used 25% of the dataset, focusing on different regions (e.g., the Americas).
C. Model updates were shared across nodes, and the global model improved through federated averaging, ensuring it could detect fraudulent chargeback patterns across multiple airlines.

This design pattern allowed the fraud detection model to continuously improve across fraud types and regions without compromising the security or privacy of sensitive data. By combining transfer learning and federated learning, the fraud detection system became a comprehensive and flexible solution for detecting multiple fraud types, while maintaining privacy and adapting to diverse data sources across the airline industry.

We conducted several experiments using a single dataset in a single-node setup, simulating a traditional machine learning environment as a comparison to a federated learning setup. The results indicate a clear improvement in performance when using federated learning. Below are the key performance metrics for each fraud detection type, comparing the single-node setup with the federated learning setup.

Single-node setup performance

In the single-node setup, the system performed well across different fraud detection types, achieving high accuracy, precision, recall, and F1 scores. Table 1 provides a detailed summary of performance metrics for each fraud type:

table 1:Key performance metrics in a single-node setup.

a) Stolen Credit Card Transactions: The system achieved 95% accuracy, 92% precision, and 93% recall, with an F1 score of 92.5. These results show that the model is highly effective in identifying anomalies in purchase patterns and device usage.
b) Chargeback Abuse: With an accuracy of 98%, precision of 97%, recall of 95%, and an F1 score of 96, BERT proved to be highly effective in detecting fraudulent chargeback disputes. This demonstrates the model’s ability to catch subtle discrepancies during transactions.
c) Loyalty Program Fraud: The model showed a high recall of 94%, indicating its strong capability to accurately detect unusual point accumulations and redemptions in loyalty programs. With 93% accuracy and an F1 score of 92.5, the system handles this type of fraud well.
d) Ticket Scams: The system’s accuracy of 96%, precision of 94%, and recall of 96%, with an F1 score of 95, reflect its strong ability to identify suspicious booking patterns, with bot detection performing especially well in this area.

Federated learning performance

After incorporating federated learning, the system was tested across two nodes and 12 communication rounds, leading to even better performance. Table 2 highlights the improvements seen in this setup:

table 2:Key performance metrics in a federated learning setup.

A. Stolen Credit Card Transactions: Federated learning increased the accuracy to 96%, precision to 93%, and recall to 94%, with an F1 score of 93.5. The improvement in anomaly detection is due to the collaborative learning approach, which allows the system to learn from multiple nodes.
B. Chargeback Abuse: The accuracy improved to 99%, precision to 98%, and recall to 96%, resulting in an F1 score of 97. This demonstrates how federated learning optimizes BERT’s NLP capabilities for more efficient chargeback fraud detection.
C. Loyalty Program Fraud: Federated learning further improved recall to 95%, indicating an enhanced ability to detect complex fraud patterns. The accuracy increased to 94%, and the F1 score to 93.5, reflecting overall better performance in this area.
D. Ticket Scams: With federated learning, the system’s accuracy reached 97%, precision 95.3%, recall 97%, and an F1 score of 96. The collaborative approach helped improve bot detection and scam identification across different channels, making fraud detection more robust.

These results underscore the advantages of federated learning, which enhances the system’s ability to generalize across diverse datasets. This approach not only improves fraud detection performance but also ensures greater data security and privacy.

This research introduces the development of an advanced fraud detection system tailored for the airline industry, utilizing a combination of transfer learning and federated learning. The system integrates TabNet for handling structured transactional data and BERT for Natural Language Processing (NLP), enabling the detection of multiple fraud types, including stolen credit card transactions, chargeback abuse, loyalty program fraud, and ticket scams.

Before federated learning was applied, transfer learning was employed to fine-tune pre-trained models, specifically using BERTBase, uncased, on airline-specific datasets. TabNet was responsible for detecting anomalies in transactional data, while BERT analyzed customer disputes and complaints to identify fraudulent patterns. The system achieved impressive performance, with 95% accuracy for stolen credit card fraud, 98% for chargeback abuse, 93% for loyalty program fraud, and 96% for ticket scams. After implementing federated learning across multiple airlines, the system improved further by enabling model training without centralizing sensitive data. Each airline trained the model locally, preserving privacy while enhancing the system’s ability to detect fraud across different environments. The results showed an overall increase in accuracy to 96%, precision to 94.5%, and recall to 95.5%, demonstrating that federated learning significantly bolstered the model’s performance.

In conclusion, the system we have built is highly effective detects with in detecting a wide range of fraud types in the airline industry. By leveraging both transfer learning and federated learning, the system offers a scalable and robust approach for real-world fraud detection implementations.

The use of advanced machine learning techniques, particularly transfer learning and federated learning, provides significant advantages for improving fraud detection in the airline industry. Transfer learning enables the adaptation of pre-trained models like TabNet and BERT, allowing the system to leverage existing knowledge and apply it to specific fraud scenarios within aviation. Federated learning further strengthens this approach by enabling collaboration between airlines while maintaining strict data privacy, ensuring that sensitive customer information remains protected.

The prototype developed in this study successfully detects various types of fraud, including stolen credit card transactions, chargeback abuse, loyalty program fraud, and ticket scams. By integrating both TabNet for structured data and BERT for unstructured text analysis, the system demonstrates the capability to handle multiple forms of fraud with high accuracy and precision. Its ability to continuously evolve and improve as new data is introduced makes it a scalable and robust solution tailored to the dynamic needs of the airline industry.

In conclusion, this study underscores the transformative potential of machine learning-driven fraud detection systems. By offering strong performance metrics, privacy-preserving features, and adaptability, the proposed system presents a forward-looking solution that can be expanded beyond airlines to other industries dealing with large-scale, sensitive data.

1. IATA (2024) Iata Iata.
2. Deloitte (2024) Deloitte's Website on Aviation Industry Insights.
3. Aras MT, Güvensan MA (2021) Challenges and key points for fraud detection in aviation. IEEE International Conference on Innovations in Intelligent Systems and Applications, Kocaeli, Turkey, pp. 20-21.
4. Prathim C, Madhuri CB, Hasitha NC, Teja NKS, Greeshma M, et. al. (2024) Strengthening airline communication network security: A zero-trust model with blockchain-based identity storage and advanced algorithms. IEEE 9^th International Conference on Science, Technology, Engineering, and Mathematics, Chennai, India, pp. 1-6.
5. Zhou X, Li W, Zhong L (2024) A supervised privacy preservation transaction system for aviation business. Peer-to-Peer Networking and Applications (17): 1898-1913.
6. Zhuang Z, Lin K, Zhang H, Chan PW (2024) Detection of turbulence anomalies using a symbolic classifier algorithm in airborne quick access record (qar) data analysis. Advances in Atmospheric Sciences (41): 1438-1449.
7. Albalawi E, M TR, Thakur A, Kumar VV, Gupta M, et. al. (2024) Brain tumor classification using MRI images: A federated learning-based approach leveraging vgg16 architecture and transfer learning. BMC Medical Imaging, pp. 1-15.
8. Qi AT, Wu F, Wu C, He L, Huang Y, et. al. (2023) Differentially private knowledge transfer for federated learning. Nature Communications 14(1): 3785.
9. Guo W, Zhuang F, Zhang X, Tong Y, Dong J (2024) A comprehensive survey of federated transfer learning: Challenges, methods, and applications. Frontiers of Computer Science 18(6): 186356.
10. Selvakanmani S, Devi GD, Rekha V, Jeyalakshmi J (2024) Privacy-preserving breast cancer classification: A federated transfer learning approach. Journal of Imaging Informatics in Medicine (37): 1488-1504.
11. Majeed U, Hassan SS, Hong CS (2021) Cross-silo model-based secure federated transfer learning for flow-based traffic classification. IEEE International Conference on Information Networking (ICOIN), Jeju Island, Korea, pp. 588-593.
12. Wang M, Lei Y, Liang Y, Lv X, Mo W (2021) AGI-fedavg: A uav power line inspection algorithm based on federated learning. IEEE International Conference on Civil Aviation Safety and Information Technology (ICCACIT), Changsha, China, pp. 20-40.
13. ML Group (2020) Credit Card Fraud Detection.
14. WBO Data (2024) World Bank Open Data.
15. Prasad B (2021) Customer Churn Prediction.
16. Openml (2014) Airlines.
17. BOT Statistics (2023) Bureau of Transportation Statistics.