Inspiration
The rapid rise of digital payment and UPI fraud showed us a critical gap in cybercrime response: investigators often lose precious time trying to guess where stolen money will be withdrawn. We wanted to turn early fraud reports into immediate, location-based intelligence that cyber cells can act on within minutes.
What it does
Our app predicts the top three ATM locations where stolen money is most likely to be withdrawn using inputs such as victim latitude/longitude, fraud ID, crime type, and amount stolen. It also assigns a risk score to each ATM based on historical fraud patterns, helping investigators prioritize where to act first.
How we built it
We developed a web interface using HTML, CSS, and JavaScript and connected it to a machine learning backend hosted on Databricks. A Linear Regression model trained on historical fraud withdrawal data learns geospatial patterns between victim locations and ATM usage. The system processes inputs from the frontend, runs predictions in the backend notebook environment, and returns the most probable ATM locations along with risk scores.
Challenges we ran into
Handling highly imbalanced fraud data and extracting meaningful patterns Mapping geospatial relationships between victims and ATM locations Designing a reliable risk score that reflects real-world fraud trends Integrating the frontend smoothly with the Databricks-hosted model for real-time results
Accomplishments that we're proud of
Converting raw fraud data into actionable geographic predictions Building an end-to-end system from model training to web deployment Creating a practical tool that can genuinely assist cybercrime investigators Achieving fast prediction times suitable for real-time response
What we learned
How geospatial data can be leveraged in fraud detection Practical deployment of ML models through a web interface Importance of risk scoring in decision-making systems Real-world challenges of working with fraud datasets
What's next for cybersecurity.ai
Incorporating live ATM transaction feeds for dynamic predictions Enhancing the model with advanced deep learning and clustering methods Adding map-based visualization for easier interpretation Expanding the system to support multiple types of financial fraud scenarios
Built With
- css
- databricksapp
- html
- javascript
- linearregression
