# Towards Adversarial Malware Detection: Lessons Learned from PDF-based Attacks

**Authors:** Davide Maiorca, Battista Biggio, Giorgio Giacinto  
**arXiv:** [1811.00830](https://arxiv.org/abs/1811.00830)  
**Date:** 2018-11-02  
**Summary page:** https://www.summarizepaper.com/en/paper/1811.00830v3/

## Key points

- Malware is a constant threat to cybersecurity, often spread through infection vectors such as documents
- Machine-learning algorithms have been effective in detecting threats, but an arms race in adversarial settings has challenged these systems
- This paper focuses on PDF-based attacks and provides a taxonomy of different approaches used to generate PDF malware and corresponding learning-based detection systems
- The authors categorize threats specifically targeted against learning-based PDF malware detectors using a framework in adversarial machine learning
- The framework identifies known vulnerabilities and novel attacks that may threaten these systems along with potential defense mechanisms that can mitigate their impact
- Promising research directions towards designing robust malware detectors in adversarial settings are highlighted.

## Summary

The cybersecurity landscape is constantly threatened by malware which is often spread through infection vectors such as documents. These vectors hide malicious code within them making it easier for attackers to use social engineering techniques to infect machines. While machine-learning algorithms have proven effective in detecting such threats, the existence of an arms race in adversarial settings has challenged these systems. This paper focuses on PDF-based attacks as a representative case of this arms race and provides a comprehensive taxonomy of the different approaches used to generate PDF malware and corresponding learning-based detection systems. The authors also categorize threats specifically targeted against learning-based PDF malware detectors using a framework in adversarial machine learning. This framework identifies known vulnerabilities and novel attacks that may threaten these systems along with potential defense mechanisms that can mitigate their impact. The findings highlight promising research directions towards designing robust malware detectors in adversarial settings which will be crucial in protecting against future cyber-attacks.