Ad fraud has been a big threat to the digital advertising market. Ad fraud has cost advertisers a whopping US$ 6.5 billion in 2017, according to the US National Advertisers Association. A recent Juniper Research study paints an even more bleak picture, predicting marketers will lose US$ 19 billion next year to fraudulent activities. This number, which reflects online and mobile ads, will continue to grow and hit US$ 44 billion by 2022 https://www.statista.com/statistics/677466/digital-ad-fraud-cost/
The industry has invested significant resources on finding successful ways to reduce the impact of ad fraud. It should be noted though, that with cyber theft or financial crimes, there is no way to eliminate the issue altogether: you can only expect to remain a step ahead of the bad guys.
Many ad fraud countermeasures have concentrated on rules-based approaches, and they are powerful means of countering basic ad fraud. However, the efforts at ad theft are getting more subtle, and conventional countermeasures today are ineffective.
As ad fraud techniques become more complex and hard to detect, so should our fraud detection systems develop in parallel and use of artificial intelligence (AI) is the only way to do that.
An AI-based ad fraud detection system actually begins with a rule-based approach as the basis but develops layers of protection by self-learning that learn from each suspicious behavior it detects. The AI-based model also has the advantage that trends can be interpreted in a lot more ways than a conventional framework.
Usually conventional rule-based models evaluate behavior across between one and three dimensions. An AI-based model analyzes more than 80 dimensions at a time so it can identify highly complex patterns of ad fraud. AI based models will evolve with self-learning as ad fraud trends develop to evade conventional systems.
To show the advantages of an AI-based strategy, we analyzed data from Januart to August this year on its own network over four months, involving over 4 billion campaign data points including ad clicks and app installs. What we found was that the fraud detection model based on AI was able to recognize twice as many fraudulent transactions as the conventional model based on the rules. The AI-based model has also proven more cost-effective for advertisers, generating a 3.6 percent higher return on advertising spending (ROAS) relative to the conventional model.
Our AI's greatest benefit, however, was its ability to spot complex, unreported ad fraud trends. What we call "the chameleon" is the pattern that our AI program marked. This is where deceptive publishers at first disguise themselves as legitimate publishers, only to later produce fraudulent downloads.
Another unusual activity that our AI detects is what we call "inventory explosion." With this trend, in the absence of an acceptable level of in-app registration operation a fraudulent publisher can produce an abnormally high inventory count.