Appier, the artificial intelligence (AI) company, today shared the results of a real-world study that it conducted on its own network to demonstrate the ability of an AI-based model to fight ad fraud. Ad fraud refers to impressions, clicks or installs that are disguised as legitimate traffic.
Based on data over four months from May to August 2017 involving over 4 billion campaign data points including ad clicks and app installs, the report showed that an AI-based fraud detection model was able to identify twice as many fraudulent transactions as a traditional rule-based model.
The advantage of an AI-based model is its ability to pick out ad fraud patterns that are difficult for traditional models to detect. One example is what Appier calls “the chameleon”, where dishonest publishers appear to be legitimate publishers at first, only to generate fraudulent installs at a later date. Another suspicious activity detected by Appier’s AI technology is termed the “inventory burst”. With this method, a fraudulent publisher generates an abnormally high inventory count in the absence of an appropriate level of in-app registration activity.
“Ad fraud has become a major threat to the online advertising industry and is projected to cost advertisers billions of dollars,” said Joe Su, CTO, Appier. “Traditional rule-based methods of detecting and mitigating ad fraud has its limitations. Appier believes an AI-based model is far more effective and we are seeing the benefits of that approach in just four months of analysis from our network.”
A traditional rule-based model typically only looks at one to three dimensions and works on rules that are programmed by humans. Appier's AI model works with simple and known ad fraud patterns, examining data along over 80 dimensions. It is also self-learning and thus able to identify suspicious patterns that it has not seen before.
“Just as with cyber fraud or financial fraud, ad fraud is becoming more sophisticated and is constantly evolving, so it’s important to be able to quickly identify and mitigate new threats. A traditional rule-based approach just cannot keep up with the fraudsters; an AI-based model is required to stay on top of evolving ad fraud patterns,” said Su.
According to Appier, there are many types of ad fraud, including:
Bot/bot network: Code that generate browsing traffic, clicks and even app installs to divert revenue to criminal sources.
Click injections: Mobile malware sends fraudulent click reports during the installation process to gain credit for the installs.
Cookie stuffing: Publishers may use 1x1 pixel ads to trigger an affiliate tracking link to steal attributions, diverting revenue.
Domain spoofing: Bad sites or traffic which appear to be premium sites.
Ghost sites: Genuine sites with content usually stolen from other sites to attract traffic. As traffic grows, these sites can appear as high-traffic sites that qualify for more lucrative advertising.
Intrusive ads: Publishers force users to visit ads to generate clicks, leading to unsatisfactory user experiences.
Invisible ads: Sites or apps which position ads in places where the users cannot actually see them to generate fake impressions. Examples are ads that are 1x1 pixel in size, or ads stacked underneath other ads.
Location fraud: Ad traffic labeled with inaccurate location information. This could be caused by intentional mislabelling at the ad network/site level, or by users who happen to be on virtual private networks (VPNs).