Introduction

Event : Seattle - Unity - QA summit - NOV6-9 2017 Hack project : Using machine learning to detect IAP fraud Team : Jeff, Mohammed, Joe, Vikas

Problem Definition

Player acquires paid game content (ex. conis, gems) without paying money. If that goes undetected then the purchases are recorded as real. The end result is a discrepancy between reported reveniew and actual revenue.

IAP Hacking Tools

• Most, not all, require rooted/jailbreak

  • LuckyPatch: Achieve fraud by changing the apk file of apps, mainly the AndroidManifest.XML file

  • Creehack (does not require root): Achieve fraud by hijacking the process and completes the purchase with its inbuilt card.

  • Cydia and Localiapstore (IOS): Cydia jailbreak device as downgrade OS possibilities, Localiapstore bypass the AppStore directly to download/install page

• Potential data we can use:

  • OS version
  • Location – currency consistency
  • App billing format
  • The app/game type (server based or not)

IAP Fraud Prevention

• Implement local and/or remote receipt validation
• Implement revenue verification
• Try to detect fraud using data

Note

We do not have access to labeled transaction data to build a real ML model so we decided to build 2 hello world demos (Octave and Python) for the sake of sharing with the team.

How ML Works? (in a min)

• Given data samples (ex. purchases)
• Predict properties of unknown data. How?
• Build a model from data matrix (do not ask us about the math please)
• Features are the columns
• Samples are the rows
• Result is continuous (regression) or discreet (classification)
• In case of IAP it is going to be valid or fraud
• Supervised: (classicfication ex. OCR or regression ex. house price)
• Unsupervised: no target values, discover similar groups (clustering)
• Training set vs testing set to evaluae the algorithm (cross validation)
• Use scikit-learn Python ML library
• Use different estimators (or models ex. SVM)
• Choosing the parameters of the model (not covered here)

Octave

• Think of it like the free Matlab
• Good for prototyping

Scikit-learn

• Open source ML lib in python
• Data standardization and normalization
• Supervised Learning Estimators (Linear reg., SVM, Naive Bayes, KNN)
• Unsupervised Clustering
• Cross-Validation
• Feature extraction and selection
• A lot more visit: http://scikit-learn.org

ML in a container

• The demo is checked to github
• Clone it, follow the instructions
• You should have an ML env template ready in 5 mins

Clone the repo

git clone https://github.com/maburrub/seattle-qa-summit-hack-ml.git

Build docker image

cd Python docker build . -t seattle-qa-summit-hack

Launch container

docker run -p 4444:7777 -d seattle-qa-summit-hack

Open browser and have fun

http://127.0.0.1:4444/demo/api/v1.0/predict?p1=3.0&p2=2.0&p3=1.0

Thanks all