The future of adblocking: stealth ad recognition by computer vision

A research team of Stanford and Princeton claims to have developed a fundamentally new ad blocking concept. The key novelties are stealth functioning and visual recognition of ads. Those are called to end the "arms race" between ad blockers and publishers who use anti-adblocking solutions.

The scientists suggest hiding an adblocker from the site its user visits. They engage a technique that malware uses to cover itself from anti-virus software. As they put it in the whitepaper, it demands "to enumerate all the entry points in the browser API that the publisher code can use to interrogate the state of the page, and to fake the responses to each such query", so that the website will "think" it displays correctly. The other technique implies creating two copies of a web page: the first has no ads and is shown to the user, the second is the one publisher’s code interacts with.

As for ad detection — modern adblockers perform it by looking through a web page source code for keywords and expressions. They are constantly changed by ad tech providers, forcing developers to create new filters and lists. The researchers suggest using computer vision for detecting ads the way a human being does: by seeing visual and behavioral elements like texts "Sponsored" or "Close ad".

Sponsored Post

This strategy relies upon the statutory requirement to label ads. Not only signatures will be detected, but images as well. Besides an adblocker will be able to emulate user behavior like a mouse click or hover over.

A Chrome extension has been created as a proof-of-concept (implementing the concept in part, excluding the technology of hiding the ad blocker by using duplicate web pages) and tested on 50 websites that use anti-adblocking solutions. It has successfully found ads on all the sites.

Andrey Meshkov, co-founder and CTO of AdGuard considers implementing of these techniques a matter of quite a distant future. Currently, this proof-of-concept extension only illustrates the idea itself. It searches through a library of known ad markers, like the AdChoice’s rectangle marker, which allows the user to customize advertising targeting in accordance with their preferences and see more relevant ads, which are marked with this very triangle.

Publishers will be able to easily bypass such recognition. And to fully implement this concept we will need an educated neural network to detect ads like a human does, by seeing typical features and attributes. Initially, this technology might have many false positive detections. It is likely that we will need different neural networks for recognizing ads in different language segments of the Internet, as the differences in ads around the world lie in more than just languages — what they look like is defined by mentality, culture, regional specifics of design and websites usability. In other words, such "training" of neural networks and combating the effects of under- and retraining will require manual work of the community and professionals, just as it now required for the creation of filters (perhaps, it will be more time-consuming).

But the ideas suggested are very interesting and perspective. Ad blockers will definitely start to detect ads using computer vision, but it’ll take years to develop the technologies and to implement them in customer products. And hardly will it "put an end to the arms race", it will rather be yet another additional strategy of ad blocking that is not likely promising to defeat ads altogether.