DitDetector

Background

DitDetector leverages bimodal learning based on deceptive images and text for macro malware detection. Specifically, we extract preview images of documents based on an image export SDK of Oracle and extract textual information from preview images based on an open-source OCR engine. And bimodal model of DitDetector contains a visual encoder, a textual encoder, and a forward neural network, which learns based on the joint representation of the two encoders' outputs.

DitDetector can detect the malicious Microsoft Office documents with deceptive information, but cannot detect the malicious MS documents that only contains VBA codes.

Install

Oracle Outside in Technology (OIT) - a tool statically export the preview images from the MS Office documents, download and related information in the Oracle blog. We have offered OIT at version of 8.5.4 in directory ./src/oit_linux_8.5.4. Before next step, please make demo:

cd ./src/oit_linux.8.5.4 && bash makedemo.sh  && sudo chmod 777 sdk/demo/exsimple

Tesseract - an open source text recognition (OCR) engine, we install it on CentOS

yum install epel-release -y
yum install tesseract -y

Create the environment through Anaconda

conda env create -f environment.yaml

You should change the prefix of conda environment in the yaml file.

unzip the test data: The test data contains malicious samples, so we have compressed the test data by zip encryption. Before running DitDetector, please use the password ditdetector to unzip the ms_files.zip archive.

cd ./data/test_data && unzip -P ditdetector -x ms_files.zip -d ./

Code Structure

The proposed detector code is in ./src, some pre-trained models and test data are in ./data.

Modules:

• Textual encoder ./src/TextCNN
• Visual encoder ./src/MobileNetV3
• Bimodal detector ./src/BiModal

End-to-end tool:

• DitDetector ./src/DitDetector

Usage

This is an end-to-end detector, which take input as the Microsoft (MS) Office documents and output a label if the sample is malicious or benign (i.e., '1' for malice and '0' for benign).

cd ./src/DitDetector
python ditDetector.py  -m ur_ms_files_dir  -e ur_export_img_dir

The ur_ms_files_dir is the directory to store the MS documents to detect, the default value is ./data/test_data/ms_files/. The ur_export_img_dir is the directory to cache the exported images by Oracle Outside in Technology (OIT), and the default value is ./data/test_data/export_img/.

Note: If you encounter the error message shown below when parsing the MS Office files with the OIT, use the export GDFONTPATH=your_font_folder_path to specify to the font folder.

EXRunExport() failed: No valid fonts found (0x0B03)

The workflow of DitDetector is as follows.

• Step 1: export the preview images of MS Office documents in ur_ms_files_dir by exsimple in Oracle OIT, and cache the exported images into the directory ur_export_img_dir.
• Step 2: extract the textual information from the preview images via Tesseract, an open-source Optical Character Recognition (OCR) tool, clean the textual information and save them into a temp file ./data/test_data/docText.csv.
• Step 3: generate the visual and textual feature representations via encoders (i.e., TextCNN and MobileNetV3).
• Step 4: concatenate these two modalities of feature representations as the bi-modal representations and feed them into the biModal to detect.
• Step 5: output the final decision whether the input document is malicious (i.e., 1) or benign (i.e., 0).

Test Environment

• Intel(R) Xeon(R) CPU E5-2630 v4 (2.20GHz and 40 cores)
• NVIDIA Tesla P100 (12GB)
• CentOS 7.9.2009
• Python 3.6.13
• conda 4.14.0

Execution Time Estimate

DitDetector execution time is flexible. In fact, the most time-consuming part is the preprocessing module, including OIT exporting preview images and OCR extracting text. Especially the OCR module, the more text in the document preview images, the more time OCR takes. In addition, we give time estimates containing the loading time of each model, and the actual sample inference time should be shorter. We detect 20 MS office documents with an end-to-end execution time of about 130s, and if we exclude both OCR and OIT modules, the remaining process takes only 13.71s.