File Type Classification based on Data Fragments from Compressed Archives for Digital Forensics

This paper presents a novel deep learning approach to classifying file types based on file fragments from compressed archive files for digital forensics such as analyzing IoT traffic. Existing solutions do not classify files in compressed archives. The proposed DL model implements and optimizes a modified Conv2D Inception network with two sets of filters FGA and FGB combined with kernel density estimation (KDE). The results of our experiment demonstrate that the modified Inception network with filter FGA and KDE using Gaussian and Epanechnikov kernels achieves a classification accuracy of $\mathbf{7 2 . 5 4 \%}$ and $\mathbf{7 3 . 1 8 \%}$, respectively. The Epanechnikov kernel with filter FGB improves the system performance and attains an accuracy of $\mathbf{7 5 . 2 4 \%}$.