Like A Human: How AI Ad Blocking Is Changing the Game

ai ad blocking

StopAd launched its AI-based ad blocking that detects advertisements like a human being. It is even able to distinguish so-called native ads that are intentionally imitating the website’s content format and style.

AI is a new electricity. In the 20th century every new product was presented as a “product + electricity” like the electric can opener. But now is the time to add intelligence to the mix.

So, Why Did We Build Our AI?

There has always been a constant conflict between ad blockers and advertising technology—and ad blocking doesn’t always win.

Ad blockers have millions of users. The most popular ad blockers scan web pages for common advertising URLs, scripts, and markup codes then block the ads. Active users help by documenting ad links and ad platforms’ domains in an open source database, Easylist. All existing ad blockers are using Easylist to provide their services.

But adtech has a money on their side.

The advertising industry spent approximately $223 billion this year to devise more ad types. Due to the high speed of this “evolution,” ad blockers become less and less effective. Additionally, ad blockers struggle with YouTube ads, Facebook sponsored posts, and native ads that mimic the layout and structure of the website under the guise of normal content.
The internet community and open source databases can’t keep up with the pace of adtech industry.

That’s why we have created our AI—to go through millions of web pages and identify ads like a human—regardless of their placement, types, sizes or whether “sponsored” label or other ad marking exists.

What’s Under the Hood of Our AI?

As a result of our research and development, we have been able to combine machine learning and computer vision. Our computer vision is built on convolutional neural networks based on InceptionV3 architecture. Our approach has resulted in a technology with high accuracy — 0.18% false positive rate. We intentionally abandoned OCR & NLP technologies. Only teaching AI to recognize the ad marking (adchoices, “sponsored by” label…) would be a short-term solution to the problem. Adtech is developing, and an ad blocking system should be more intuitive and flexible.

Our approach has resulted in a technology with high accuracy — 0.18% false positive rate.

How We Train Our AI?

  1. Our bots collect web pages as reports (one bot can scan 60K web pages per day). We plan to scale to check 1M websites per day.
  2. Collected data sets are used for further AI training. The system divides pages into logical blocks, which are defined as advertising, non-advertising, and native advertising then sends them to a neural network.
  3. After the learning process, AI presents its results. It examines each rendered page by processing logical blocks. As a result, for each page, it concludes whether it contains ads.
  4. After verification of accuracy, a blocking rule is automatically generated and sent to the StopAd rules database.

How Accurate Is Our AI?

You can check our AI ad blocking by yourself with the help of our API. Here is how:

1) Find your favorite website, filled with ads. For example,

2) Take a screenshot of the banner ad with any application that allows you to upload it to the web. Try to capture the banner itself (without extra parts of the screen) or block of ads (if there is a native ad). As a result, you should get a link to an image.
For example:
Banner ad screenshot:

Native ad screenshot:

3) Send it to our AI through an API.{url_to_image} ,
where {url_to_image} is a link to your screenshot.
For example:

4) Get the answer.

After following the previous link we saw the answer:

So what does this all mean? Here’s the key:

“noAds”—the possibility that this image is not advertising (from 0 to 1),

“banner”—the possibility that this image is an banner ad (from 0 to 1),

“native”—the possibility that this image is an native ad (from 0 to 1).

As you can see, the AI returns a .99 of 1.0 classification that the test we submitted was a banner ad, which it certainly is. Give it a try. Let us know how we do!