A deep learning identification method of a raw hexadecimal byte sequence

Abstract

The application provides a deep learning identification method for a raw hexadecimal byte sequence, based on a JSANet-BC model, including the following steps: standardizing and vectorizing a metadata raw byte stream, and then converting discrete bytes into continuous representations containing sequential information through an embedding layer and position coding; using a Transformer encoder block to learn self-attention features in the byte dimension; further refining and down-sampling features through a one-dimensional convolution module containing a channel attention mechanism, and finally outputting an identification result through global feature aggregation and a classifier. The above method can realize an identification approach that does not depend on display metadata analysis and can automatically learn robust features from raw data.

Images