# EvilModel: Hiding Malware Inside of Neural Network Models

**Authors:** Zhi Wang, Chaoge Liu, Xiang Cui  
**arXiv:** [2107.08590](https://arxiv.org/abs/2107.08590)  
**Date:** 2021-07-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.08590v1/

## Key points

- Advanced malware campaigns require covert delivery and evasion of detection
- "EvilModel" introduces a method of embedding malware into neural network models
- Neural networks lack explainability but have strong generalization capabilities, making them ideal for embedding malware
- Malicious code can be strategically embedded into neurons without significantly impacting model performance
- Up to 36.9MB of malware could be embedded into a 178MB-AlexNet model with only a 1% loss in accuracy
- Leveraging neural networks for malicious purposes is a growing trend in cyberattacks
- The study highlights the risks associated with neural network-assisted attacks and offers insights for enhancing defense mechanisms

## Summary

In the realm of advanced malware campaigns, the ability to deliver malicious software covertly and evade detection is paramount. In a recent study titled "EvilModel: Hiding Malware Inside of Neural Network Models," authors Zhi Wang, Chaoge Liu, and Xiang Cui introduce a novel method for achieving this goal using neural network models. These models are known for their lack of explainability and strong generalization capabilities, making them an ideal vehicle for embedding malware without significantly impacting their performance. By strategically embedding malware into the neurons of neural networks, the researchers demonstrate that malicious code can be delivered covertly while maintaining the structural integrity of the model. This approach allows malware to evade detection by traditional antivirus engines, as the overall architecture of the neural network remains unchanged. The authors conducted experiments showing that up to 36.9MB of malware could be embedded into a 178MB-AlexNet model with only a 1% loss in accuracy, all without triggering any suspicions from antivirus engines like VirusTotal. As artificial intelligence continues to proliferate across various industries, leveraging neural networks for malicious purposes is becoming an increasingly prevalent trend in cyberattacks. The findings presented in this paper shed light on the potential risks associated with neural network-assisted attacks and offer valuable insights for enhancing defense mechanisms against such threats. By exploring innovative techniques like those proposed in "EvilModel," cybersecurity professionals can better prepare themselves to combat evolving forms of malware that exploit advanced technologies like neural networks.