Abstract

Although the core concept inside an deep learning AI model is common, currently there is no uniformed standard for representing AI models. Each popular DL framework, such as TensorFlow and Caffe2, uses its own formats. In addition, vendors, such as XiaoMi, Qualcomm, Tencent, all come up their own model formats. Most of DL framework and run-time systems make some sanity check to the AI models before loading them from applications. However, the properties being checked are ad-hoc. A framework or its run-time environment can load an AI model with defects or containing malformed neural networks.

This talk presents a static checker that find defects in AI models. The checker can detect flaws such as inconsistency of data dimensions, topology errors in neural networks. The checker also highlights the vendor specific content in AI models. We show that, such defects, if not detected, can cause damages to the run-time environments and cause security risks to deep learning applications. The danger to be shown in this presentation includes system hangs, crashes, and leak of critical application and user information. These flaws affect DL frameworks from well-known mobile vendors, and DL systems that support cloud-based AI services.