SOC: TNG

Slides

https://www.lvh.io/DesdemonaIntro

Introduction

Who am I?

Rackspace

‌

Not a sales pitch, but…

"Rackspace's" SOC

RMS: GA in Jan 2016

No SOC before that?!

CSOC & ISOC

Different mechanics

Different services

• Lots of in-house tooling
• Endpoint & phishing focus

This talk: CSOC

Challenges

Lots of APIs!

Don't reinvent the wheel

(I don't want to implement FIM)

Opinionated set of services

Vertical expansion

(vs core competency)

Vendor cycling

Conclusion

Need to integrate with lots of APIs, fast

Scale

Security-, not SOC-as-a-Service

Fire hose

Multi-platform

Lots of OSes

Lots of operators

Lots of platforms

Big bets

Research project

(If we knew what we were doing…)

Motivation

What makes a SOC?

Correlation

SOC analysts

Close cooperation

Dev team ⟷ SOC

RTT hours instead of months

Better correlation tools

• E.g. query language
• Cheaper experiments
• More insights

Education

Enough programming/ML/data analysis

Hosting provider advantage

Economies of scale

Already know the environment

"Internal" data

e.g. from shared network infra

Storage

Storage is pretty cheap

Store by default

• Paper trails
• ML training data
• Experimental detection

Software

Lisp!

Clojure(Script)

Why does it matter?

• I like magical Lispy power
• Rapid development cycle
• Less boilerplate

Ex: enrichments as pure functions

Arbitrary choice computation vs storage

Ex: query language

Cross-event correlation

ip(x) = ip(y)
and (port(x) = {http,https}
     or x in my-case)
and suspicious(y)

Query language

• Compiles to logic program
• infix ⇒ sexps ⇒ minikanren ⇒ JVM/JS
• ~identical code on JVM & JS
• Next goal: index-assisted search

Hypothesis: 3 OOM I care about

• <1s: feature detection
• <1h: analysis over larger dataset
• >1d: training ML models

Onyx

• Stream processing
• Like Storm (but better, I think)

Cool Onyx feature: Just functions

Easy to re-use code if experiment fails

Pipeline branches

• Modify pipelines on the fly
• Cheap experiments without jeopardizing prod
• "Would this have caught the attack from last Friday?"

Machine learning

Example: DGA

E.g. Conficker.C, Murofet, BankPatch, Bonnana, Bobax

Generated domains

cikiugcaqcegsimg.org

tosecureonweestablishment.com

Offense

Every day:

• Generate 50k domains
• C&C registers 1 domain
• Box connects to 500

Defense: prevention

• Reactionary
• Lots of domains

Defense: blacklist

Works, but:

• Reactionary
• Regexen 😩
• Not even regexen 😭

Defense: machine learning

Proxy:

• human can detect
• small {state, feature} space

Defense: machine learning

H₀: legit domain ~ other legit domains

Training set (tagged)

• Legitimate domains (Alexa 1M, OpenDNS)
• Malicious domains (Data Driven Security)
• Malicious domains from botnets (conficker, pushdo, ramdo, rovnix, tinba and zeus)

Features

• Char-level features
• Shannon entropy
• Word-level features

Char-level features

• Shared character sequences
• "Can I pronounce this?"

Shannon entropy

Language-level features

• Google N-gram corpus
• Top 100k words

Language-level features

What words are there?

cikiugcaqcegsimg.org

tosecureonweestablishment.com

• Are these (sort-of) words?
• How many? How long?

Language-level features

Do these words make sense?

bobsfishtacklebait.com

tosecureonweestablishment.com

Refer to n-grams

Methodology

• 80/20 train/eval
• Resample for balance
• Gradient boosted trees
• Evaluate precision & recall

Precision

How many identified items are relevant?

Recall

How many relevant items are selected?

Confusion matrix

Some languages are hard

Another example: flux domains

compromised hosts ⇒ routing mesh

Features

• TTL
• Set of IPs at FQDN
• Latency

Results

2-3 nines on the main diagonal

Conclusion re: ML

Great results with not a lot of effort

Further efforts

• Other ML algorithms
• Better features
  • More language corpora
  • IP allocations

Back-end hardware

Barreleye

Barreleye

Barreleye

Barreleye

• POWER8 @ 3.7-4.1 GHz, 192 HW threads
• >200MiB on-die cache
• 2TB RAM, >200GiB/s mem bandwidth
• CAPI: direct access to flash, FPGA

Barreleye (just one more)

Client hardware

Desktops have more RAM

64GB per desktop

⇓

who cares if Chrome's RSS is 60GB?

Same code on client & back-end

When in doubt, cheat

Cheating: GeoIP

GeoLite2-City-Blocks*

200MB CSV, 33MB mmdb

Cheating: optimizing JS compilers

First pass: self-hosting CLJS

Second pass (cheating)

Metron

Differences

Partial coverage

Questions?

https://www.lvh.io/DesdemonaIntro