Neural Parameter-Space Classification And Applications to Hardware Explanation

Abstract

Neural networks encode information into the parameters they learn during training. This thesis explores the use of machine learning techniques both to differentiate a single trained neural network from other similar networks and to identify the dataset used to train it. We apply these techniques in the context of computer hardware reverse engineering, where we identify unknown, “black box” computer peripherals by modeling their observed input/output behavior with memory-based deep recurrent neural Networks (DRNNs). Once trained, these networks encode important information about the original device. We present a large dataset of trained neural networks that mimic the behavior of simple computer peripherals, and explore the differences in encoded parameters to surface identifying features of these devices. While less practical at scale for our chosen context, the underlying experiments and observations into classifying neural networks presented here are broadly applicable to the larger field of neural network explanation.