“Any fool can tell a crisis when it arrives. The real service to the state is to detect it in embryo.”
— Isaac Asimov, Foundation
Until recently, the approach to fighting mobile ad fraud has been defensive and reactive.
This cat-and-mouse game has continued unabated with the bad actors continually evolving, varying their means of attack and using techniques that have grown in complexity and scope over time. The Methbot operation showed how the ad fraud industry has become more sophisticated and organized.
The methods used by ad fraud crime syndicate are fast changing in the wake of increasing adoption of advanced data analytics and machine learning algorithms by adtech. After all, an automated learning system that continuously detects and prevents fraud is much more desirable and more efficient over static rule-based and manual intervention systems.
At InMobi, we are taking control of the battleground by using the attackers’ own predictable yet ever-evolving methodologies to not just detect fraud as it happens, but to actively prevent it as well.
The primary strategy used to counter fraud today is to define known frauds, and then to focus on detecting those anomalies with pre-determined fixed thresholds. This is ineffective and inefficient against constantly evolving, newer fraud methods for a simple reason — how do you find something when you don’t know what to look for?
Thus, a framework that can detect such anomalies sooner, i.e., what is genuine and what isn’t, and neutralize them quickly is essential.
The ideal framework constantly learns and memorizes what’s normal. This way when something anomalous is detected, no matter how sophisticated, it can react.
InMobi’s Deep Learning algorithms adaptively learn complex patterns and extrapolate to detect newer types of fraud.
Our deep learning algorithms automatically model every device and user on the network, enabling the system to ‘memorize’ how everything from requests to post-click events normally flow. It extrapolates this data in real time to detect fraud.
Anomaly detection techniques and unsupervised learning methods are employed to discover deviant behaviors, across multiple dimensions, that are indicative of fraud. In addition to this, fraud indicators are tracked over time to determine, with higher certainty than ever before, repeated fraud activity from particular devices or market cuts. Further, time-series analysis enables us to spot fraudulent on-or-off patterns over time and better catch serial fraudsters as well as better filter 'accidentally labeled' fraudsters.
There’s no substitute to human intelligence. Therefore, in addition to using deep learning, InMobi’s industry leading framework uses a three-pronged approach supported by data, insights and human intervention.
Deep learning algorithms function better with more data and better data. Thus, the availability of quality data across regions and verticals is a critical component to building frameworks that effectively use machine learning to detect fraud more reliably and accurately.
InMobi, having been part of the incredible mobile growth story for over a decade, has vast amounts of quality data worldwide — including data specific to fraud. This helps not only in training our models iteratively and quickly, but also to validate its accuracy and reliability (i.e, measure bias/variance) on test data.
Lastly, in addition to data, having knowledge and experience in fighting ad fraud is key to developing better algorithms and models.
As an added note, this framework is enabled by default on the InMobi ad network to all advertisers and has been extensively tested over the years. To know more, visit inmobi.com/trust.