TL;DR
This paper introduces FiFTy, a novel deep learning approach for identifying file types from raw data fragments, crucial for digital forensics and data recovery. FiFTy utilizes a compact convolutional neural network (CNN) architecture that learns features directly from byte sequences, significantly outperforming previous methods in both accuracy and speed, and is released as open-source along with a new, large dataset.
What Problem Does This Paper Address?
In digital forensics and data carving, identifying the type of a file fragment is a fundamental but challenging task. Traditional methods rely on hand-crafted features (like magic bytes or statistical properties) which are often brittle, struggle with a large number of file types, and are computationally expensive due to the feature extraction step. This bottleneck limits their scalability and efficiency, especially when dealing with fragmented data or memory dumps. The authors aim to overcome these limitations by developing a more robust, accurate, and significantly faster file-type identification system.
Methods
The core of FiFTy is a 1D convolutional neural network (CNN) architecture. Instead of relying on pre-defined features, FiFTy takes raw byte blocks as input and uses a trainable embedding layer to represent each byte. This embedding is then processed through convolutional blocks (convolution and max-pooling layers) to gradually reduce dimensionality and capture spatial patterns. Finally, average pooling, dropout, and dense layers are used to map the learned features to class probabilities. The authors also employed the Tree-structured Parzen Estimator (TPE) for hyperparameter optimization, systematically searching for the best network configurations across various scenarios.
Datasets
The authors introduce a novel dataset named FFT-75, which comprises 75 diverse file types, making it the largest and most balanced dataset reported for this task to date. This dataset was created by sampling blocks from publicly available files. For evaluation, they also used the GovDocs dataset and a DFRWS challenge dataset.
Key Insights and Ideas
- End-to-end learning from raw bytes: FiFTy's success hinges on its ability to learn relevant features directly from byte sequences, bypassing the need for manual feature engineering. This is a significant departure from prior work and a key enabler of its performance.
- Trainable embedding space: Using a trainable embedding layer for bytes, similar to word embeddings in NLP, allows the model to capture nuanced relationships between byte patterns, leading to better generalization.
- Compact CNN architecture: The proposed CNN architecture is designed for efficiency, balancing accuracy with computational speed, which is critical for real-world forensic applications.
- Hyperparameter optimization is crucial: The extensive use of TPE highlights the importance of systematically exploring the hyperparameter space to achieve optimal performance for different scenarios and block sizes.
- Dataset diversity matters: The creation of the FFT-75 dataset with 75 file types, including many multimedia and RAW photo formats, was essential for evaluating the model's performance on a realistic and challenging set of data.
Key Results
FiFTy demonstrated superior performance compared to state-of-the-art methods like Sceadan, NN-GF, and NN-CO.
- Accuracy: Achieved an average accuracy of 77.5% on the 75-file-type dataset with 4096-byte blocks, significantly outperforming Sceadan (69.0%).
- Speed: Processed data at approximately 38 sec/GB, which is over an order of magnitude faster than Sceadan (9 min/GB).
- Scalability: The model's architecture and training process are designed to scale to a large number of file types.
- Scenario-specific performance: Accuracy improved as the number of classes decreased, reaching up to 99.7% in a specialized JPEG vs. others scenario.
- Misclassification analysis: While FiFTy performed well, common misclassifications were observed for HEIC/MOV (due to similar encoding) and for archive formats (due to high entropy).
Limitations and Future Work
- Dataset limitations: The authors acknowledge that the FFT-75 dataset has some dependency issues, particularly with files containing embedded objects (like images within PDFs or PPTs), which can confuse classifiers. Future work could involve more sophisticated data labeling or carving techniques for such compound file types.
- Model complexity: The hyperparameter optimization sometimes led to complex models. Future work could explore the trade-off between model complexity, computational effort, and accuracy, and investigate more advanced architecture search methods.
- Generalization to diverse content: While FiFTy performs well on common file types, the authors suggest that additional care might be needed for specific file types (e.g., video files with varying codecs) to accommodate diverse content characteristics. They also propose exploring random field models for block-level agreement and incorporating known header/footer information.
Summary updated