Technology
December 03, 2025
Fraud detection: how AI redefined financial security
Fraudsters adapt quickly, which requires continuous evolution in protection systems
Financial fraud continues to be one of the biggest challenges in the digital economy. In 2022 alone, losses exceeded 1.2 billion euros in the United Kingdom and 8.8 billion dollars in the United States (Khalid et al., 2024). This highly dynamic environment requires financial institutions to move away from reactive actions and adopt increasingly sophisticated predictive approaches.
The digitalization brought efficiency, but also expanded the opportunities for fraud. Technologies like Chip & PIN reduced in-person fraud, but stimulated the migration of attempts to the online environment, especially in Card Not Present (CNP) transactions (Edge & Sampaio, 2009). In this scenario, fraudsters adapt quickly, which requires continuous evolution in detection systems.
Historically, fraud detection relied on fixed rules, like “if X happens, block”. These were simple and predictable systems, easily bypassed. The first major shift occurred with machine learning, which began to learn patterns directly from data, allowing the identification of behaviors outside the norm (Carcillo et al., 2021).
At this stage, two groups of techniques emerged: (1) supervised models, trained with transactions labeled as legitimate or fraudulent, and (2) unsupervised models, capable of detecting anomalies without relying on labels, especially useful for novel frauds.
This stage also consolidated the importance of feature engineering. Behavioral variables such as sum of expenses, transaction frequency, temporal variations, and patterns by locality significantly expanded the capacity of models to capture subtle deviations (Seera et al., 2024).
Deep learning in action
With the popularization of neural networks, a new technological leap emerged. While the first architectures analyzed isolated transactions, models like LSTMs began to capture the sequential behavior of the user, identifying patterns over time (Roseline et al., 2022; Jurgovsky et al., 2018).
This type of analysis contextualizes the transaction. A purchase may seem legitimate in isolation, but deviate from the client’s recent pattern, and LSTMs capture exactly that.
Beyond the transactional level, neural networks have also been applied to identify fraud risk at the organizational level, with models capable of analyzing internal and external factors and achieving accuracy greater than 90% (Krambia-Kapardis et al., 2010).
Graphs and advanced detection
The current frontier of fraud detection is dominated by Graph Neural Networks (GNNs). Unlike models that analyze individual users only, GNNs consider the entire ecosystem as a connected network. Customers, devices, IP addresses, merchants, and transactions form a dynamic graph.
This approach allows for the identification of fraud rings, suspicious connections between accounts, coordinated patterns among multiple users, and recurrent use of fraudulent devices. Graph-based models have been delivering superior results in detecting organized and sophisticated fraud (Chen et al., 2024; Zhu et al., 2021).
Challenges
Despite the advances, implementing AI in production brings concrete challenges. The first is the strong data imbalance. Fraud represents only 0.172% to 0.36% of transactions in large databases (Carcillo et al., 2021; Dornadula & Geetha, 2019). This requires specific techniques, such as resampling by oversampling or undersampling, to prevent the model from learning to classify everything as “legitimate”.
Another critical point is measuring success. Traditional metrics, such as accuracy, are ineffective in this context. The financial sector relies mainly on Recall, to avoid letting fraud escape, on Precision, to avoid blocking legitimate customers, on F1-Score and AUC-PR, which are suitable for imbalanced datasets (Kim et al., 2019).
In practice, detection involves hybrid systems. Automatic models perform real-time transaction scoring, while a human team analyzes only the most critical cases. Furthermore, ensembles that combine multiple models increase the stability and accuracy of the results (Khalid et al., 2024; Randhawa et al., 2018).
The Future of prevention
Two emerging trends point to the next generation of financial security.
| Digital twins: a virtual replica of each customer’s behavior allows simulating a transaction before authorizing it. If the digital twin identifies a significant deviation, the operation can be blocked preventively (Chatterjee et al., 2024). |
| Blockchain and federated learning: financial institutions can train models collaboratively, but without sharing sensitive data. Blockchain technology provides a secure and transparent ledger, while federated learning allows for training distributed models while preserving privacy (Chatterjee et al., 2024). This combination addresses a historical bottleneck in the sector: improving models without compromising confidentiality. |
The detection of fraud has evolved from simple rule-based systems to solutions that integrate advanced AI, specialized neural networks, and graph structures. Nevertheless, the challenge remains in constant mutation, and the effectiveness of institutions will depend on the ability to integrate new approaches, continuously update models, and combine automation with human analysis.
In the end, security is not a destination, but a continuous process of adaptation. And, in the technological race against financial crime, the winner is the one who evolves non-stop.
Who wrote this column
José Erasmo Silva
