FiFTy: Large-scale File Fragment Type Identification using Neural Networks

Govind Mittal, Pawel Korus, Nasir D. MemonarXiv 2019 · OpenReview

Abstract

We present FiFTy, a modern file type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space, akin to successful natural language processing models. Our approach dispenses with explicit feature extraction which is a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 file types—the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy, and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of approx 38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool, Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset are available publicly online.

Introduction

In digital forensics, data carving is the critical process of recovering files from storage devices when file-system metadata is missing, corrupted, or deliberately obfuscated. Investigators must identify the type of raw data fragments—often as small as 512 or 4,096 bytes—without any external context. Historically, this task relied on hand-crafted statistical features such as byte frequency, n-grams, or Shannon entropy. However, calculating these features is computationally intensive, creating a significant bottleneck in high-volume forensic analysis. Furthermore, legacy tools struggle to differentiate between modern, high-entropy multimedia formats like compressed video or RAW images.

The FiFTy Architecture

To address these limitations, we developed FiFTy, a compact 1-D Convolutional Neural Network (CNN) that learns directly from raw bytes. By treating file fragments as sequences of "words" (bytes), FiFTy leverages techniques from Natural Language Processing (NLP).

The FiFTy neural network architecture, featuring a trainable embedding layer followed by convolutional and pooling blocks.

The core innovation lies in the trainable byte embedding layer. Instead of using sparse one-hot encodings or expensive bigram matrices, each byte value (0-255) is mapped to a dense, real-valued vector space. This allows the network to learn the "context" of byte transitions automatically. These embeddings are then processed through 1-D convolutional blocks and max-pooling layers to extract characteristic patterns, followed by dense layers for final classification.

Datasets and Evaluation

We introduced the FFT-75 dataset, the most diverse and balanced corpus for file fragment identification to date. It contains over 100,000 blocks for each of 75 file types, with a strong emphasis on modern multimedia formats (including 11 RAW photo and 7 video formats). We also validated FiFTy on the public GovDocs corpus and the DFRWS 2006 forensic challenge image to ensure generalizability and real-world applicability.

Performance and Results

FiFTy sets a new state-of-the-art for both accuracy and throughput. On the challenging 75-class task with 4,096-byte blocks, FiFTy achieved 77.5% accuracy, significantly outperforming the previous leader, Sceadan (69.0%).

Comparative analysis of FiFTy against legacy file-type identification tools across various datasets and metrics.

The speed improvements are even more dramatic. FiFTy processes data at approximately 38 seconds per GB (0.64 min/GB), whereas Sceadan requires 9 minutes per GB under identical conditions. This represents a 14x speedup for 4,096-byte blocks and up to a 34x speedup for 512-byte blocks. When specialized for specific tasks, such as "JPEG vs. Everything Else," FiFTy reaches an exceptional 99.7% accuracy.

Conclusion and Future Work

By eliminating the feature-extraction bottleneck and utilizing trainable embeddings, FiFTy provides a scalable solution for modern forensic challenges. While the model currently faces difficulties with compound files (e.g., JPEGs embedded within PDFs), future work will explore using conditional random fields to analyze neighboring blocks, potentially improving consensus and recovery for larger, fragmented files.

Summary updated

Abstract

We present FiFTy, a modern file type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space, akin to successful natural language processing models. Our approach dispenses with explicit feature extraction which is a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 file types - the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of approx 38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool - Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset are available publicly online.
BibTeX
@misc{mittal2019fifty,
  author = {Govind Mittal and Pawel Korus and Nasir D. Memon},
  title = {{FiFTy: Large-scale File Fragment Type Identification using Neural Networks}},
  year = {2019},
  eprint = {1908.06148},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.06148},
}

Loading related content...

Figure