Every second, millions of online banking transactions flow through the world’s banks, and fraudsters try to hide among them. BankNet is an AI framework that analyses transactions in real time to catch fraud as it happens.
The problem
Internet banking has grown rapidly, and with it the need for systems that make secure decisions in real time. Banks must spot fraud instantly without slowing down legitimate customers.
What the research did
We built BankNet, a predictive analytics framework that combines big data tools with a bidirectional LSTM neural network, a type of AI that reads sequences of events in both directions to understand patterns over time, such as a customer’s transaction history.
What we found
BankNet detected fraud with 98.5% accuracy and achieved a root mean squared error of 0.0159 in its predictions. It handled data rates of up to 1,000 Mbps with minimal delay.
Who benefits
- Bank customers, who are better protected from fraud
- Banks and fintech companies needing fast, scalable decision support
- Regulators concerned with trust in digital finance
Limits and what comes next
Fraud patterns change constantly, so models need regular retraining. Real deployment also requires careful attention to customer privacy and to avoiding unfair false flags.
The team
This research was led by Kaushik Sathupadi (Google) with Sandesh Achar, Shinoy Vengaramkode Bhaskaran, me and Jia Uddin (Woosong University).
Frequently asked questions
How does AI detect banking fraud?
It learns patterns of normal and fraudulent transactions and flags unusual activity in real time.
How accurate is BankNet?
It detected fraud with 98.5% accuracy in testing.
Can it handle a big bank’s traffic?
It handled data rates up to 1,000 Mbps with minimal delay in experiments.