Inspiration
Financial institutions' growing adoption of blockchain faces significant challenges: regulatory uncertainty, data privacy concerns, and the need for secure collaboration. Inspired by the need to balance transparency with privacy, we aimed to develop a solution that enables financial institutions to detect and mitigate risks in blockchain transactions without exposing sensitive user data.
What it does
Our solution is built to ensure the safe sharing of information without compromising data privacy across financial institutions and an anomaly detection model that can flag our suspicious transactions.
How we built it
The ‘preprocess_transactions’ function transforms raw transaction data into numerical features suitable for machine learning. It extracts time-based features from transaction timestamps and applies one-hot encoding to categorical features, such as transaction type or country. This ensures consistent data formatting for both historical data used in training and new transactions for anomaly detection.
The ‘train_anomaly_model’ function utilises the Isolation Forest algorithm to learn patterns of normal transaction behaviour from past data. Isolation Forest is capable of effectively identifying outliers without required labeled data, making it suitable for fraud detection. The model is trained on processed historical data, and an encoder is returned alongside the model for consistent feature transformation.
The ‘detect_anomalies’ function applies the trained Isolation Forest model to new transactions to predict which transactions are anomalous. Detected anomalies are logged, and details of the suspicious transactions are saved into a SQLite database. This is done using the ‘save_anomalies_to_db’ function. After anomaly detection, non-anomalous transactions are appended to the historical dataset. The model is then retrained with the updated data to adapt to new transaction patterns. This continuous learning loop allows the system to learn and evolve over time, improving the accuracy of anomaly detection for a more robust system.
Challenges we ran into
Blockchain integration is difficult to implement as we had limited knowledge on before starting this project, however, we were able to overcome the challenges through research and trial and error in our code.
Accomplishments that we're proud of
We managed to successfully integrate Federated Learning, ZKPs and anomaly detection model into a functional prototype. Furthermore, our prototype is feasible and can be implemented across financial institutions after further fine-tuning.
What we learned
We learnt the technical intricacies of Federated Learning and how it can be applied in decentralised systems, as well as the importance of blockchain as a tamper-proof ledger for collaboration.
What's next for Live Laugh Code
We would like Transition from a centralised anomaly detection model to a federated learning framework across multiple institutions for enhanced privacy and scalability, and optimise AI models for performance, scalability, and security.
Log in or sign up for Devpost to join the conversation.