SQRRLTHREAT HUNTING PLATFORM
ADAM FUCHS
CTO, SQRRL
COMMITTER, ACCUMULO
MEMBER, ASF
© 2016 Sqrrl Data, Inc. All rights reserved.
2
Accelerating Investigations
© 2016 Sqrrl Data, Inc. All rights reserved.
3
The SqrrlThreat Hunting Platform
SECURITY DATA
NETWORK DATA
ENDPOINT/IDENTITY DATA
Firewall / IDS Threat
Intel
Bro
SIEM
Alerts
NetflowProxy
ProcessesHR
© 2016 Sqrrl Data, Inc. All rights reserved.
4
Sqrrl Architecture
Visualization + API
Physical
Data Storage
Data Model
Processing
Interface
Audit
Encryption
Labeling+Policy
Query Engine:
Accumulo Iterators
Bulk/Graph Processing:
YARN + Spark
Raw Events Linked Data
HDFS Accumulo+
Commodity Hardware
© 2016 Sqrrl Data, Inc. All rights reserved.
5
The Apache Accumulo Project
Accumulo Stores Sorted Key,Value Pairs
High Performance Writes
Great Scalability
Embedded Processing (Iterators)
We leverage Accumulo for:
Low-Latency Information Retrieval
Indexing
Distributed Processing
GraphOrganization
Ingest-Time Aggregation
Secure Storage
Behavioral Analytics
© 2016 Sqrrl Data, Inc. All rights reserved.
7
Attack Chain Behavior detection
Adversary behavior is modeled based on a kill chain
Kill chain alignment of behavior detection analytics:
Helps to determine attack penetration and risk
Supports arguments of completeness of detection coverage
© 2016 Sqrrl Data, Inc. All rights reserved.
8
Kill Chain-Based Behavioral Analytic Example
• Lateral
Movement:
Multiple host
logins,
credential
theft
• Active
Directory
• Windows
event logs
• Unsupervised
machine
learning for
rarity
detection
• Graph
algorithm for
chaining
• Analyst
whitelisting
of false
positives
© 2016 Sqrrl Data, Inc. All rights reserved.
9
Collating Results ForVisualization and Analysis
Behavioral Analytics Entity Risk Scoring
Raw Data
Modeled Data (Graph)
API Applications
AnalyticsAnalyticsAnalyticsAnalytics
Target. Hunt. Disrupt.

SQRRL threat hunting platform

  • 1.
    SQRRLTHREAT HUNTING PLATFORM ADAMFUCHS CTO, SQRRL COMMITTER, ACCUMULO MEMBER, ASF
  • 2.
    © 2016 SqrrlData, Inc. All rights reserved. 2 Accelerating Investigations
  • 3.
    © 2016 SqrrlData, Inc. All rights reserved. 3 The SqrrlThreat Hunting Platform SECURITY DATA NETWORK DATA ENDPOINT/IDENTITY DATA Firewall / IDS Threat Intel Bro SIEM Alerts NetflowProxy ProcessesHR
  • 4.
    © 2016 SqrrlData, Inc. All rights reserved. 4 Sqrrl Architecture Visualization + API Physical Data Storage Data Model Processing Interface Audit Encryption Labeling+Policy Query Engine: Accumulo Iterators Bulk/Graph Processing: YARN + Spark Raw Events Linked Data HDFS Accumulo+ Commodity Hardware
  • 5.
    © 2016 SqrrlData, Inc. All rights reserved. 5 The Apache Accumulo Project Accumulo Stores Sorted Key,Value Pairs High Performance Writes Great Scalability Embedded Processing (Iterators) We leverage Accumulo for: Low-Latency Information Retrieval Indexing Distributed Processing GraphOrganization Ingest-Time Aggregation Secure Storage
  • 6.
  • 7.
    © 2016 SqrrlData, Inc. All rights reserved. 7 Attack Chain Behavior detection Adversary behavior is modeled based on a kill chain Kill chain alignment of behavior detection analytics: Helps to determine attack penetration and risk Supports arguments of completeness of detection coverage
  • 8.
    © 2016 SqrrlData, Inc. All rights reserved. 8 Kill Chain-Based Behavioral Analytic Example • Lateral Movement: Multiple host logins, credential theft • Active Directory • Windows event logs • Unsupervised machine learning for rarity detection • Graph algorithm for chaining • Analyst whitelisting of false positives
  • 9.
    © 2016 SqrrlData, Inc. All rights reserved. 9 Collating Results ForVisualization and Analysis Behavioral Analytics Entity Risk Scoring Raw Data Modeled Data (Graph) API Applications AnalyticsAnalyticsAnalyticsAnalytics
  • 10.

Editor's Notes

  • #9 Identify Pain Point It is hard to find attackers moving from a beechhead machine to more interesting machines in a sea of login data Every Windows machine has tens to hundreds of logins a day normally Project datasets Use a subset of windows event login data most likely to contain attacker movement Constrain Output LMs will only be of certain login chain shapes Localized in time Identify Algorithms Use modern classifiers on projected log event set to identify logins that are more or less likely to be in a LM Use motif search algorithms to find patterns of logins that fit the output constraints Self-learning and feedback