Gaussian low-pass channel attention convolution network for RF fingerprinting

Summary: This paper rigorously investigates the Frequency Principle in the training dynamics of Deep Neural Networks (DNNs) at three stages: initial, intermediate, and final. The results, presented in the form of theorems, are general and applicable to multilayer networks with general activation functions and a wide range of loss functions, providing a theoretical foundation for a better understanding of the training process of DNNs.