# Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal

**Authors:** Xinwei Liu, Jian Liu, Yang Bai, Jindong Gu, Tao Chen, Xiaojun Jia, Xiaochun Cao  
**arXiv:** [2207.08178](https://arxiv.org/abs/2207.08178)  
**Date:** 2022-07-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.08178v1/

## Key points

- The paper addresses the issue of visible watermark removal by deep neural networks (DNNs)
- Proposes a defense mechanism called Watermark Vaccine
- Visible watermarks are commonly used to protect copyrights of digital images
- DNNs can remove watermarks without damaging the host images
- Adversarial machine learning approach proposed for defending against watermark removal
- Two types of vaccines introduced: Disrupting Watermark Vaccine (DWV) and Inerasable Watermark Vaccine (IWV)
- DWV induces damage to both host image and watermark after passing through watermark-removal networks
- IWV aims to keep the watermark intact and noticeable even after attempts at removal
- Extensive experiments conducted to evaluate effectiveness of DWV/IWV on various networks
- Results demonstrate successful mitigation of copyright infringement risk by making it difficult for DNNs to remove watermarks from digital images
- Valuable contribution in addressing vulnerability of visible watermarks to DNN-based attacks
- Offers proactive defense strategy for protecting image ownership and copyrights in digitized world.

## Summary

The paper titled "Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal" addresses the issue of visible watermark removal by deep neural networks (DNNs) and proposes a novel defense mechanism called Watermark Vaccine. Visible watermarks are commonly used to protect copyrights of digital images, but recent research has shown that DNNs can remove these watermarks without damaging the host images. To counter this threat, an adversarial machine learning approach is proposed for defending against watermark removal. The authors introduce two types of vaccines as part of their defense mechanism: Disrupting Watermark Vaccine (DWV) and Inerasable Watermark Vaccine (IWV). DWV induces damage to both the host image and its watermark after passing through watermark-removal networks while IWV aims to keep the watermark intact and noticeable even after attempts at removal. Extensive experiments were conducted to evaluate the effectiveness of DWV/IWV in preventing watermark removal on various networks. The results demonstrate that the proposed defense mechanism successfully mitigates the risk of copyright infringement by making it difficult for DNNs to remove watermarks from digital images. This paper presents a valuable contribution in addressing the vulnerability of visible watermarks to DNN-based attacks and offers a proactive defense strategy that can help protect image ownership and copyrights in an increasingly digitized world.