The Problem: Medical AI needs to be portable (for phones) but also secure against errors. The Solution: I built a Quantized MobileNetV2 (73% smaller than standard) and tested it against adversarial attacks.
Key Findings:
Accuracy: 94% on Test Data.
Efficiency: Compressed from 9MB to 2.5MB.
Security Flaw: A noise injection of just 0.05 caused the AI to misdiagnose a tumor as healthy.
Code Structure:
NeuroScan: Vulnerability Analysis of Edge-Deployed Medical AI.ipynb: The training and attack experiments.
app.py: A live demo of the lightweight model.