top of page
mascon-icon.png
mascon.png

Hugo Guillaume

Security engineer
Konvu
greg-fawson-foto.1024x1024 Kopie.jpg

Hugo is a security engineer at Konvu, building AI-driven vulnerability triage systems that separate real exploitability from scanner noise. He also lectures on security module for some French engineering school, delivering hands-on coursework to graduate students.

Previously, Hugo spent two years at the French Ministry of Defense in vulnerability research and reverse engineering.

When "Encrypted" Isn't: Protecting On-Device ML Models in Mobile Apps

Session

Ship a model to a phone and you've shipped it to whoever owns the phone. Teams know this, so they encrypt the model on disk and move on. That's not protection. It's a speed bump.

I'll show why, with an SDK I took apart and disclosed: Cambridge Mobile Telematics' SensorFlow, which rides inside insurance apps like Intact Ajusto, belairdirect Automerit, and State Farm Drive Safe & Save. Inside is an encrypted PyTorch crash classifier. Its verdict feeds premium pricing and claims, so the model is worth hiding, and the encryption is the whole defense.

I pulled the cleartext model out without rooting a device and without Frida. The method is almost dull, which is the point: emulate the SDK's own native library, libsensorflow.so, in software with a Unicorn harness, and let the app's decryption routine hand me the plaintext in emulator memory. The app unlocks its own secret.

bottom of page