Exploring 1D Data Augmentation Techniques for Improved File Fragment Classification

File Fragment Classification (FFC) is essential for digital forensics, facilitating file type identification without relying on metadata or intact headers. Despite advances in model architectures for FFC, the potential of data augmentation for improving model robustness and generalization has not been extensively explored, especially for 1D byte sequence data. This study investigates the applicability and impact of two well-known image-based augmentation techniques, Masking and CutMix, on 1D byte-sequence-based FFC models. We conducted comparative experiments applying Masking, CutMix, and the previously proposed Gaussian Bit Flip (GBFlip) augmentation across several model architectures, including FiFTy, DSCNN, ByteRCNN, ResNet1D, and XMP. Experimental results reveal that augmentation effectiveness varies depending on model architecture. Notably, both FragMask and FragMix yielded performance gains in deeper models such as XMP and ResNet1D, across both short and long fragments.