Fraud Detection in Crypto Transactions
Scenario
In the dynamic and complex world of cryptocurrencies, the ability to identify potentially fraudulent transactions and accounts has become a crucial priority. Leveraging its unique expertise in Mathematical Intelligence, Deix collaborated with Deep4IT to develop an innovative solution capable of analyzing blockchain data and classifying accounts based on their fraud risk.
The challenge
Deep4IT, active in the development of products for financial data analysis, faced the need to enhance its fraud detection capabilities in the crypto domain. The main challenge was to build a system capable of classifying accounts as fraudulent or legitimate based solely on transaction history, without relying on external data sources. This required identifying anomalous behavioral patterns and effectively managing the irregular and often incomplete nature of blockchain data.
The solution
Deix addressed this challenge by developing a custom code library, integrable into Deep4IT’s systems, for training a classification model and performing inference on new accounts. The key element of this solution is the application of transformer-type neural networks—state-of-the-art architectures in the field of natural language processing—to the analysis of transaction time series.
Refined handling of temporal data
The model can effectively manage gaps and temporal misalignments inherent to blockchain transactions.
Extraction of complex patterns
The transformer architecture makes it possible to identify long-term correlations and dependencies within transaction sequences, revealing fraud patterns that are difficult to detect with traditional approaches.
Continuous learning
Continual Learning methods allow the model to be constantly updated.
Results achieved
The models developed by Deix demonstrated superior performance compared to baseline models (CNN and LSTM), achieving an accuracy of over 80%. This translates into a significant reduction in false positives and an improvement in the efficiency of fraud detection systems.
Future developments
The technology developed in this project has broad application potential in various fields:
The algorithms can be adapted to analyze time series from traditional bank accounts, online transactions, and other types of financial data.