Abstract

The paper discusses about reversing the deep learning AI model revealing its architecture, critical hyperparametser that can be exploited by malicious actor. This reversing is beyond finding password, keys and buffer overflows. The paper will discuss detail analysis of reversing the model from different models viz. googlenet, llama etc and various formats such as hd5, onnx and bin. The parameters discussed after reversing are related to tensors in deep learning models viz. sparsity of matrix, architectural flow, weights and biases that are fundamentals to any AI model. Also for language models the tokanizer reversing through model will be discussed. In short mathematical structure of deep learning model will be reversed.