Towards Improving the Reliability of Deployed Deep Learning Software

Deep learning makes mobile apps smarter, but on-device DL models are vulnerable to theft. My research shows that attackers can reverse-engineer these models to steal their details. To protect them, I first developed two methods: static model obfuscation, which hides key model representation, and dynamic model obfuscation, which confuses attackers at runtime. Additionally, I created CustomDLCoder to extract and hide essential parts of the model. These techniques help keep DL models secure on mobile devices, protecting user data and app integrity.