How to Successfully Harness AI to Combat Fraud and Abuse - RSA 2018

[@sebastiengoutal6213]

Thanks for this very interesting talk Elie. You mention data augmentation technics to improve the classifier : I know this is often used in computer vision or malware detection, as one can easily add noise/geometric transformation to a picture or add noise to a malware - but do you know if it can be used to generate targeted email attacks such as business email compromise or CEO fraud? This is a field where it is extremely difficult to collect data. Thanks! Sebastien

[@Victor-em1ms]

17:05 Regarding what you called 'single class classification', which is basically unsupervised learning. Most people will use something like PCA to find odd behavior, but such apporach in feature reduction to spot outliners require good ammouts of feature to begin with. And the issue of PCA, is that it ruins human interpretation. You advise for low feature problem to get as much information as possible from the context: IP adress, screen resolution... But my question is at which point can you tell wether or not those informations are usefull, or if there just polluting your model? Cause at the end your deconnected from the feature. You have a good performing model, but if you could not use some feature and keep the accuracy, you would gain in computing power. But how to tell which feature to drop with your apporach? Do you have pratices to indentify the not so usefull feature early on? Or just do you do just trial & error?

[@junijjang]

Thanks for the nice contents. It's really helpful :)

[@xiaodongsu1376]

Great talk, Elie. Could pls send me a copy of the slides? thanks

[@xiaodongsu1376]

Cooool

[@ebursztein]

elie.net/talk/how-to-successfully-harness-ai-to-combat-fraud-and-abuse/

[@felipeonl]

Is it possible to use cache injection in a browser game?

[@ebursztein]

If its online there are likely server side checks