展开查看详情
1 .How to Use Millions of
Mobile Activity Logs to
Understand Our
Customers, in Real Time!
Aaron Colcord, FIS Global
Kevin Mellott, FIS Global
#DevSAIS18
2 . Who is
FIS Global?
• One of the largest global
FinTech companies
• Customers are banks and credit unions
• We’re FIS Digital Finance, Mobile Data
and Analytics
• Ecosystem of products and services
built around core banking
#DevSAIS18 2
3 .The Evolution of Banking
#DevSAIS18 3
4 .Digital Banking
• Check our Account Balances
• Deposit a Check
• Pay Someone/Bill
• Get Money out of the ATM
• We should be able to replace our actual wallets with
the digital wallet.
• Let’s talk about Fraud as an aspect of understanding
#DevSAIS18 4
5 .Fraud Detection Datapoints
– Frequency of Transactions
– Location of the Transactions
– Pattern
– Items purchased
– Total Spend
– Transaction Amount
#DevSAIS18 5
6 .Fraud Detection has evolved too
• The Algorithms that fraud uses have only grown.(This is
Anomaly Detection)
• The Accuracy of when something is fraud and not fraud has
only grown more accurate
• Biggest Drawback happens on the backend.
– After/During Transaction
– Emphasis has been catching faster and notifying faster
– Being Predictive about what transactions mean
– Grabbing more Detail about the Transaction to make the prediction
#DevSAIS18 6
7 .This could be better
• We are emphasizing more work faster
to close our fraud.
• When Fraud is detected, usually our
access is cut and that is good to stop
more bad stuff from occurring. But if
there is actually no fraud, that isn’t
good.
#DevSAIS18 7
8 .With Mobile, is this the best we can do?
We can prove it is you by BioMetrics
We can reasonably assume you have your phone on you
We can track what security features you have enabled
Prove how you are and where you are
Enrich the Experience, Add Behavior and Feedback Loop
#DevSAIS18 8
9 .Scenarios
ONE TWO THREE
Behavioral Is this your Weird
Pattern primary Transactional
device? Pattern
#DevSAIS18 9
10 .Traditional Fraud Detection
• Supervised learning activity
• Trained using ALL transactions
• Anomaly detection (simplified)
#DevSAIS18 10
11 .Challenges
• Building a training dataset
• User behavior is unique
• Evolve what we already know
#DevSAIS18 11
12 .Scenario One
ONE TWO THREE
Behavioral Is this your Weird
Pattern primary Transactional
device? Pattern
#DevSAIS18 12
13 .Behavioral Patterns
• Is it normal for this person?
• Personalized anomaly detection
• Indirect correlation to transaction
• Unsupervised analysis
#DevSAIS18 13
14 .Physical Behavior
• When using our mobile app:
– Wifi Network
– Longitude & Latitude
– Device Info (make/model/os)
– App Info (installId/version)
• Metadata associated with every action
#DevSAIS18 14
15 .User Behavior
• Authentication Details
– Biometric (touch, face)
– 2FA (SMS, etc)
– User/Pass
• Payment Method
• Device Used
– Have we seen it before? When?
#DevSAIS18 15
16 .Time-based Behavior
• Repeating patterns
– Pay landlord same amount each month
• Time/distance since last entry
– Multiple concurrent sessions
– Different locations, short time period
• User tendencies
– Ex: Usual to purchase at 4am local time?
#DevSAIS18 16
17 .Scenarios
ONE TWO THREE
Behavioral Is this your Weird
Pattern primary Transactional
device? Pattern
#DevSAIS18 17
18 .Device Overload
• We live in a world of devices
– Voice assistants, watches, phones, etc.
– Will keep growing with IOT
• People often use the same one for purchasing
#DevSAIS18 18
19 .Managing Devices
• Register upon first time use
– Includes step-up authentication method
– Provides a way to "activate" privileges
– Establishes a baseline
#DevSAIS18 19
20 .Scenarios
ONE TWO THREE
Behavioral Is this your Weird
Pattern primary Transactional
device? Pattern
#DevSAIS18 20
21 .Weird Transactions
• Comparing a specific transaction
– "Typical" for the merchant
– "Typical" for the customer
• Buying 20 laptops online
– Normal for a business owner
– Sketchy for an individual's account
– Establish customer profiles (clustering technique)
#DevSAIS18 21
22 .How Does it Work?
#DevSAIS18 22
23 .Architecture
#DevSAIS18 23
24 .ML Pipeline
#DevSAIS18 24
25 .Model Storage
• Re-builds in background
• Databricks Jobs API
#DevSAIS18 25
26 .Model Usage
• Loaded within Structured Streaming
• Transformation produces result
• Available to other apps
#DevSAIS18 26
27 .Message Broker
• AWS Kinesis
• Many will work
• Kafka, Azure, etc.
#DevSAIS18 27
28 .Communicating Fraud
• Results write to S3
• Event notification triggers SQS
• Queue connects to mobile app
• Requires step-up authentication to verify
#DevSAIS18 28
29 .Feedback Loop
• Fraud department activities
• Feeds into our training data
• Model regenerates via Databricks Job
• Always using latest model
#DevSAIS18 29
热门城市
