Robustness in Deep Neural Network: A Focus on Software Tampering and Counter Measures

Over the past several years, there has been a proliferation of "Deep Neural Network" (DNN) models that have been constructed as well as implemented. These models require safeguarding against potential tampering by malevolent individuals. This paper aims to explore the significance of "recoverable, self-embedding fragile watermarking approach for deep neural network DNN" models in order to safeguard the integrity of the models. This system possesses the capacity to not only identify and locate the altered parameter blocks in the framework, but also to accurately recover the compromised values. The verified data and recovery data are derived through a comprehensive analysis of the specific attributes of the DNN model that requires safeguarding. These data are then embedded into the model using a reference sharing mechanism, without compromising its original functionality. This enables the recovery of the model parameters even when subjected to various levels of tampering.