Security researcher Altay Akkus has revealed techniques for extracting artificial intelligence models from mobile applications, specifically demonstrating the process using Microsoft’s Seeing AI and Adobe Scan apps. The research, published in a technical blog post, shows how AI models intended for on-device use can be accessed and extracted using various tools and methods. Using Frida, a dynamic instrumentation toolkit, Akkus successfully extracted the currency detection model from Microsoft’s Seeing AI app, which is designed to help visually impaired users identify different types of currency.
The researcher detailed how the extraction process works by intercepting the TensorFlow Lite framework’s model loading process and dumping the unencrypted model files. The extracted model was verified using Netron, a neural network visualization tool, confirming the successful capture of the complete model including weights and biases.
The researcher notes that this technique can be applied to any mobile application using TensorFlow Lite for on-device AI processing. While demonstrating the technical feasibility of the process, Akkus acknowledges potential legal considerations, noting that AI models are typically considered intellectual property and their extraction may have legal implications.