Deep PDF Parsing to Extract Features for Detecting Embedded Malware

摘要

The number of PDF files with embedded malicious code has risen significantly in the past few years. This is due to the portability of the file format, the ways Adobe Reader recovers from corrupt PDF files, the addition of many multimedia and scripting extensions to the file format, and many format properties the malware author may use to disguise the presence of malware. Current research focuses on executable, MS Office, and HTML formats. In this paper, several features and properties of PDF Files are identified. Features are extracted using an instrumented open source PDF viewer. The feature descriptions of benign and malicious PDFs can be used to construct a machine learning model for detecting possible malware in future PDF files. The detection rate of PDF malware by current antivirus software is very low. A PDF file is easy to edit and manipulate because it is a text format, providing a low barrier to malware authors. Analyzing PDF files for malware is nonetheless difficult because of (a) the complexity of the formatting language, (b) the parsing idiosyncrasies in Adobe Reader, and (c) undocumented correction techniques employed in Adobe Reader. In May 2011, Esparza demonstrated that PDF malware could be hidden from 42 of 43 antivirus packages by combining multiple obfuscation techniques. One reason current antivirus software fails is the ease of varying byte sequences in PDF malware, thereby rendering conventional signature-based virus detection useless. The compression and encryption functions produce sequences of bytes that are each functions of multiple input bytes. As a result, padding the malware payload with some whitespace before compression/encryption can change many of the bytes in the final payload. In this study we analyzed a corpus of 2591 benign and 87 malicious PDF files. While this corpus is admittedly small, it allowed us to test a system for collecting indicators of embedded PDF malware. We will call these indicators features throughout the rest of this report. The features are extracted using an instrumented PDF viewer, and are the inputs to a prediction model that scores the likelihood of a PDF file containing malware. The prediction model is constructed from a sample of labeled data by a machine learning algorithm (specifically, decision tree ensemble learning). Preliminary experiments show that the model is able to detect half of the PDF malware in the corpus with zero false alarms. We conclude the report with suggestions for extending this work to detect a greater variety of PDF malware.