Banking Secure application - for fraud detection

A behavior-based authentication (BBA) system that uses machine learning to continuously validate the identity of a mobile banking user by analyzing real-time behavioral and contextual signals — preventing account takeovers, session hijacking, and fraud without sacrificing usability.

1. SentinelAuth: AI-Powered Continuous Behavioral Authentication for Mobile Banking SURAKSHA CYBER HACKATHON 2025

2. The Problem: Authentication Gaps Current Pain Points Real-World Threats Fraudulent transactions, session hijacking, and account takeovers are becoming more common. One-time validation is provided by passwords, OTPs, and even biometrics, which leaves sessions exposed. Conventional techniques are unable to identify behavioral deviations or post-login anomalies. SIM swap attacks give hackers the ability to evade 2FA and intercept OTPs. Complete access to bank accounts is made possible by credentials that have been stolen through phishing or data leaks. Sensitive information and session tokens are exposed by Man-in-the-Middle (MITM) attacks on public Wi-Fi. Attackers frequently take action after logging in, when security is at its weakest.

3. Introducing SentinelAuth

4. Behavioral Intelligence – What We Track Touch Patterns Tracks how the user taps and swipes: pressure, speed, rhythm, and gesture style. Everyone interacts with their screen slightly differently—these patterns are hard to fake. 2. Device Motion Analyzes how the phone is held and moved: tilt, grip angle, and stability. Useful to detect unfamiliar handling or shaky behavior under duress. 3. Navigation Flow Monitors how users move through the app: typical paths, speed, and tap patterns. Fraudsters often navigate differently than genuine users. 4. Temporal Patterns Looks at when and how long users typically use the app. Unusual times or sudden fast sessions can be warning signs. 5. Location Behavior Compares login locations with past usage: regular zones vs sudden changes. Helps flag suspicious access from unexpected places or devices.

5. Modular And Scalable Design

6. Real-Time Fraud Detection with Tiered Actions Score Range Interpretation Action 0.0–0.6 Normal Full access 0.6–0.8 Mild anomaly OTP / biometric challenge 0.8–0.9 Strong anomaly Lock sensitive features >0.9 Likely fraud Logout + alert Risk Score → Action : Autoencoder (Unsupervised) : Learns normal user behavior. Fails to reconstruct anomalies → flagged. Isolation / Random Forest : Detects rare, outlier behavior patterns. (Optional) RNN / LSTM : Models sequences of user actions for context-aware detection. Our Approach to Detection Context-Aware Adaptation Travel behavior, time-of-day, device switch, and user age are considered to reduce false positives.

7. Privacy & Compliance No unprocessed data or keystrokes are saved or transmitted to the cloud. Timing, rhythm, and other behavioral characteristics are hashed and anonymized. Complete compliance with DPDP and GDPR TensorFlow Lite for local processing Edge Case Handling Elderly : Lenient thresholds and acceptance of slow behavior Disabled users :optional assistive templates for user support Travelers :Travel mode toggle, IP/GPS-aware scoring Stressed users : Tap pressure and navigation jitter tolerated

8. Performance Scale 90% Accuracy Achieved in identifying legitimate vs. fraudulent activity. 2% False Positive Rate Minimizes legitimate user friction. 200ms Latency Real-time processing for immediate responses. 3% Battery Impact Minimal energy consumption ensures user adoption.