Finding Vulnerabilities in TensorFlow

Explore a cutting-edge fuzzing approach for detecting vulnerabilities in machine learning frameworks in this 21-minute Black Hat conference talk. Delve into the limitations of traditional API fuzzing and discover how to uncover deep vulnerabilities hidden within complex code logic. Learn about structure-aware model mutation, tensor generation and scheduling, input placeholder mutation, and op mutation techniques. Gain insights into the increasing security risks associated with rapidly developing AI frameworks like TensorFlow, PyTorch, and PaddlePaddle. Examine real-world results and takeaways from this innovative vulnerability detection method, presented by a team of six cybersecurity experts.

Intro
Two Business Models in Al systems
Structure-aware Model Mutation
A Keras Model Example
What is Tensor?
Random Tensor Generation
Input Tensor Scheduling
Input Tensor Mutation Takeaways
Input Placeholder Mutation
Op Mutation
Our Results