Coursera Spring Sale
40% Off Coursera Plus Annual!
Grab it
This conference talk explores the process of reversing deep learning AI models to reveal their underlying architecture and critical hyperparameters that could potentially be exploited by malicious actors. Delve into techniques that go beyond traditional security concerns like password extraction, key discovery, and buffer overflow detection. Learn detailed analysis methods for reversing various model types including GoogleNet and Llama, across different formats such as HD5, ONNX, and binary files. Discover how to extract fundamental AI model parameters related to tensors, including matrix sparsity, architectural flow, weights, and biases. The presentation also covers tokenizer reversing techniques for language models, ultimately providing insight into how to uncover the mathematical structure that forms the foundation of deep learning models.
