Nullcon Goa 2025: Reversing Large Deep Learning AI Models - Yashodhan Vivek Mandke

This talk addresses the complexities of reversing deep learning AI models to uncover their architecture and critical hyperparameters that may be vulnerable to exploitation. The speaker presents a detailed analysis of various models, including GoogLeNet and LLaMA, as well as formats such as HDF5, ONNX, and binary files. Key topics include the examination of tensors related to sparsity of matrices, architectural flow, weights, and biases that are fundamental to AI models. Additionally, the presentation discusses techniques for reversing tokenizers in language models, ultimately revealing the mathematical structures underlying deep learning architectures.