Robster
Rob your precious machine learning model
@austinkim
Introduction
This project was developed for detect and extract from mobile application (APK only)
Currently, only static analysis method is supported.
See the detail description in blog (Korean)
https://iam.namjun.kim/research/2021/02/24/rob-your-precious-machine-learning-model/
Features
Detection
- File signature matching
- static DEX (dalvik execution format) file method call trace
Extraction
- Extract model using file scanning (signature matching)
Phases
Engine
Engine for detect and extract the ML model.
Backend
The HTTP API server for connecting between robster-engine and end user.
Frontend
TO-DO
Supported ML libraries
- Tensorflow Lite (Tensorflow, Detect and extraction)
- Tesseract (Google, Detect only)
- Ncnn (Tencent, Detect only)
- MNN (Alibaba, Detect only)
- Caffe (Berkeley AI Research, Detect only)
Pre-requirements
- Docker
- docker-compose
How to use
Run command
docker-compose up -d
Reference
The inspiration for this project is coming from this paper.
- Zhichuang Sun, Ruimin Sun, and Long Lu, "Mind Your Weight(s): A Large-scale Study on Insufficient Machine Learning Model Protection in Mobile Apps", arXiv:2002.07687