ExpShield-demonstration

This page demonstrates how ExpShield protects text content from scraping. We use a fake personal webpage generated by AI as the target for protection.

• The source code embedded with our defense is in index.html
• This HTML file can be opened in any browser and also available at https://toan-vt.github.io/ExpShield-demo.

Defense Mechanisms

We employ two ExpShield defenses to protect paragraphs of the webpage's content:

1. Invisible Style Defense

This method involves wrapping the protected text in a tag with a specific CSS class that makes the text invisible to the human eye but still present in the HTML source code.

We insert <span class='invisible-text'>ExpShield</span> into each word. The corresponding CSS is:

/* CSS */
.invisible-text {
    font-size: 0;
    line-height: 0;
}

2. Invisible Character Defense

Insert invisible characters directly into words, e.g., Today → Tod​ay (Here an invisible character is inserted between d and a.)

---

Browser Rendering

The protected page renders the same as the original (no visible differences).

Web Scraping Results

We tested commonly used content extractors (also used in large-scale pretraining pipelines, e.g., The Pile [1]). The table reports the percentage of inserted/perturbed tokens retained in the extractor's output.

Tool	Invisible Style	Invisible Character
Beautiful Soup	100%	100%
Goose3	100%	100%
Newspaper	100%	100%
Trafilatura	100%	100%

---

Reproducibility

All scripts and instructions to reproduce the extraction and retention numbers are in the web-crawling directory of this repository.

[1] https://arxiv.org/abs/2101.00027