Fraud is costly and running rampant in online marketplaces, social networks and communities across the web. Effectively spotting and stopping fraud on a global scale requires determining the signals most closely correlated with fraudsters. Fraudsters vary in their methods, machines and machinations — but there are some commonalities and signals that point to likely fraud.
To provide a glimpse into some of the leading indicators of fraud, we aggregated more than 100 different signals across 500,000 real world browser-based devices throughout January 2016. We looked for patterns in the 10,000 (or 2%) of those devices that were in the hands of fraudsters and contrasted those with the other 98% of devices in the hands of good or organic users.
Seven interesting anomalies emerged that were leading indicators of fraud:
1. 32-bit OS running on 64 bit processors: A transaction is 8x more likely to be fraudulent if the device configuration matches this description — often because fraudsters use cracked versions of older Windows machines which are imaged and then explicitly programmed for greater control like cookie cleaning.
2. Fresh cookies: Fraudsters clear their cookies 90% of the time whereas organic users clear cookies only 10% of the time. Thus cookie age is a strong fraud signal, and unlike the baked items, browser cookies are more likely to be good the older they are.
3. Null values: There is another feature in browsers which is “Do Not Track” (http://donottrack.us/). For organic/real users the possible options are “Yes”, “No”, “Unspecified”; with “No” (70% of the times) as the default settings. With fraudsters on the other hand this value is often “null” which is not among possible organic values. There are more such browser configuration parameters where fraudulent devices have values other than the possible organic values.
4. Flushed browser referrer history: Fraudsters often flush their browser referrer history. <5% of the organic population explicitly filter their referrer history using third party plugins, extensions. Fraudsters as a population are 5x more likely to do this.
5. Bad apples dont use Macs: While Windows desktop and laptop have a dominant market share organically (90%+ overall) and 70%+ in our sampled data of users, 96%+ of fraudsters use Windows.
6. Fraudsters do not install a lot of plugins and extensions: They tend to “keep it simple” with 90% of fraudsters having less than 5 plugins in the browser. By comparison, good users have more plugins and in fact 1 in 20 of the organic population have more than 25 plugins/extensions installed — which is in a way risky and asking for trouble.
7. Fraudsters dont go incognito: But its not just the fraudsters behavior thats revealing. There is are also leading indicators for good users. For example, a user in private mode is more likely to be good than bad. Surprisingly, fraudsters do not enable private mode — in fact, organic users are 3x more likely to prefer private mode.
The global fraud landscape is always changing. But through the application of machine learning with broader and broader data sets, we are able to build a detailed map of behaviors, software, hardware and configurations that are likely indicators of fraud over time. With this digital fraud intelligence our customers are better able to efficiently and cost-effectively stop fraud without slowing down good customers.
Written by Rahul Pangam, CEO of Simility
Rahul is the Co-Founder and CEO of Simility. Being a fraud detection industry veteran, he believes in combining the power of algorithms to recognize similar and dissimilar signals with the ability for humans to create meaning and giving front-line fraud fighters tools that empower them to put their domain expertise and knowledge to use without needing to write code.