How machine learning is a ‘requisite’ for ad fraud detection

Marketers/advertisers are bundled with data today. They are collecting data behind every touchpoint the consumer makes, right from click data, install data, engagement data, etc.

In today’s world, there are 2 major activities marketers are involved in:

• Using click and install data, marketers keep investigating different forms of campaigns to drive bigger volumes down the digital funnel

• Using engagement data, marketers study channels of engagement and message throughout the lifecycle of a digital consumer to enable a higher LTV

However, this isn’t enough to study whether the incoming data is attributed correctly or not. Vanity matrices today are merely numbers on a digital dashboard, but the correctness is immensely suspected throughout.

Finding attribution manipulation can be problematic and estimating an analogical behaviour of the traffic to a constant is merely impossible. For the same, because of the problem and the largeness of the data, it requires machine led understanding of the data over time.

Usage of filters, boundary conditions, threshold, etc gives a good descriptive statistical understanding of the data in hand and can estimate rule-based anomaly finding. However, this misses on predictive and prescription data science.

AI In Advertising Fraud

In order to build a true machine learning model, one must look at the data very closely and build a homogeneous learning model that only injects consumer journey behaviour as a learning variable.

Examples of Common Ad Fraud Schemes in which ML helps :

1. Some sub-publisher based mobile application’s track consumer’s keyword search in google play store or iOS store, and if a consumer searches for a particular advertiser that is active and running performance led campaigns, a click is generated. These clicks hijack traffic from other networks and steal the organic traffic as well. A CTIT learning might not be enough to highlight such an anomaly, as these hijacks generally have a CTIT of more than 20 seconds.

2. Some sub-publisher based mobile application’s track customer’s’ APK changes. In case, a customer installs a particular android or apple app package, a click is generated to hijack this form of traffic. Generally, these lie in the CTIT anomaly limit of 20 seconds, but a back timed click is sometimes even injected to claim the attribution.

3. Incase of installs and engagement based KPI for performance campaigns, APK drops are a common thing for acquiring new customers from tier 2, tier 3 and rural India. The sub-publisher-based mobile application works as adware and takes the rights of installing new-APK on the mobile device. These kinds of installs are generally greyed at half a cent at the value to many aware marketers. Marketers opt for it for reaching the quick 1 million mark, or high listing on google play or apple store. However, these installs are generally not brand-safe and might allow data theft.

Finding attribution manipulation is not easy. The objective behind the mere click manipulation is to hijack the last-click attribution model for monetary gains. Anyone with mere contacts of these adware/malware enabled apps can help one in growing the business in no time which in return fuels corruption, fake news, tax evasion and cross-border cyber warfare.

Detection for these kinds of traffic sources is a must, as it incentivises the above, but also affects digital consumers and the country in the long run.

As published in : Financial Express