Microsoft Researchers Slash Skype Fraud By 68% 114
mask.of.sanity writes "Life could become more difficult for fraudsters on Skype thanks to new research by Microsoft boffins that promises to cut down on fake accounts across the platform. The research (PDF) combined information from diverse sources including a user's profile, activities, and social connections into a supervised machine learning environment that could automate the presently manual tasks of fraud detection. The results show the framework boosted fraud detection rates for particular account types by 68 per cent with a 5 per cent false positive rate."
That's nice. (Score:5, Informative)
So let me get this straight...
Your new filter works better than today's filter...against today's spam
But today's spam is designed to circumvent today's filter, and spammers will change their techniques as soon as you switch to the new filter.
This is the classic Antivirus problem, where new and unusual AV programs get great ratings until they become popular and virus developers start coding with them in mind.
And now you've also published how your new filter works, to make it even easier for spammers to circumvent your new filter. Great.
You see... (Score:3, Informative)
Re:BAD MATH! (Score:4, Informative)
TFS (and TFA, actually) are poorly phrased: the actual research article (the linked PDF) specifies (and I quote):
The aim of our work is to go beyond the present, sophis-ticated defenses, and to detect "stealthy" fraudulent users, namely, those that manage to fool those defenses for a relatively long period of time. Our concrete objective is to catch these stealthy fraudulent users within the first 4 months of activity. Our results indicate that, with our methods, we are able to detect 68% of these users with a 5% false positive rate; and we are able to reduce by 2:3 times the number of these users active for over 10 months.
So they didn't increase their detection rate by 68%, they increased it to 68%. And 5% false positive is pretty good: 95% confidence interval is standard in scientific research (outside things like physics which is able to achieve much much higher confidence by means of vastly larger data sets), which means a 5% false positive is exactly what you'd expect with proper scientific methodology ( based on a quick scan that seems to be exactly what they were aiming for). And of course higher false positive is actually better in the case of fraud detection than lower detection rate (since little is harmed by a false positive, while false negatives can directly result in people losing money).