Ranveer Chandra and Dina Katabi Learning Communication Rules Srikanth Kandula.

Ranveer Chandra and Dina Katabi

Learning Communication Rules

Srikanth Kandula

Network Admins. are Groping in the Dark

Focus on Traffic Volume• TCP=80%, HTTP=30%

• Adapt report categories (e.g., AutoFocus)– Much traffic from ports 500-600

But, What’s Going On?• Traffic follows plan?• Misconfigurations• Suspicious Traffic

(Active) user browsing web, reading/sending mail(Automatic) SMS scan on a network, outlook refresh

Besides focusing on volume, learn rules underlying the traffic Besides focusing on volume, learn rules underlying the traffic

• Infer the actual behavior of applications– AFS root servers direct traffic to volume servers evenly– mail to the incoming MX, is forwarded onto group MXes

• Notice misconfigurations and badness– these clients shld not be talking on known command-control ports this server shld not be responding to DHCP requests

– this mail server shld not attempt connections to non-existent MXes

flowY flowXWhenever flowy happens, flowx is likely to occur

Rule

Rule

tX X X XY Y Y

If you could learn such rules directly from a trace,If you could learn such rules directly from a trace,

(http DNS)

Report all significant rules with no specific knowledge about a trace

Mining for Rules is Hard• How to define significance?

– When is a group of flows interesting enough to report?

• Avoid observer bias but cannot evaluate everything– Focus on one server, miss what you are not looking for

• Practical, deal with noise, search quickly

eXpose1. A scoring function for significance2. Heuristics that bias search toward high hit-rate3. Empirical validation on enterprise traces

eXpose1. A scoring function for significance2. Heuristics that bias search toward high hit-rate3. Empirical validation on enterprise traces

Overview

• Packet trace to Activity Matrixo Rows are 1s windows; Columns are flows o Is flow active in [timei-1, timei )? (at least one packet)

• Association rule mining (X,Y are r.v. for columns)• Need not worry about interleaving• Dependencies are at these time-scales (an rtt, a server response)

PacketTrace

flow1 … flowK

time1

…

timeR

Activity Matrix

Rules

All windows in [.25s, 2s] range yield similar rulesAll windows in [.25s, 2s] range yield similar rules

Which Rules are Significant?• High Joint Probability?

o X, Y may occur very often individually (e.g., breeze, sun shining)

• High Conditional Probability?o Say Y occurs only when X does, but both are rare (lottery, buy a jet)

X Y

* Measures fraction of change in Y due to X

• High Joint Probability?• High Conditional Probability?• We use mutual information (combines the two)

( ) ( ) ( )( ) ( ) ( )

( )YP

XYPYXP

YP

XYPXYPYXScore

¬

¬¬+=→

|log

|log

* Trades off dependency & frequency

Score=0, if Y is independent of X

Score=Max, if Y is fully dependent on X

* Encodes Directionality KerberosReservation

Which Rules are Significant? X Y

• Negative Correlation– Flows with little overlap

Y… X …

( ) ( ) ( )( ) ( ) ( )

( )YP

XYPYXP

YP

XYPXYPYXScore

¬

¬¬+=→

|log

|log

P(Y|X) 1 leads to high score

Modifying Scores for Networking

• Negative Correlation– Flows with little overlap

• Long Running Flows– Large downloads, ssh/remote desktop– Trivial overlaps with long flow– Distinguish new vs. present– Present rules reported only if small mismatch in freq.

• Too Many Possibilities– Bias, focus on pairs with at least one common IP– Miss rules, but hit-rate up 1000x and costs down 10x

Y… …

Y… X …

( ) ( ) ( )( ) ( ) ( )

( )YP

XYPYXP

YP

XYPXYPYXScore

¬

¬¬+=→

|log

|log

X

P(Y|X) 1

Modifying Scores for Networking

Generics- Miss, if no client accesses server often+ Rules that abstract away parts of a flow

Server

Database

Client : Server Server : Database

Reservation

Kerberos

Client : Server Server : Database *

Client : Rsrv. Client : Kerberos

Client : Rsrv. Client : Kerberos * *

(any client)

(any client, but same on both sides)

To do this automatically,• what to abstract? (IP addresses at non-server port)• which pairs to consider for rule?

– flows match IP, generics match abstracted IP

To do this automatically,• what to abstract? (IP addresses at non-server port)• which pairs to consider for rule?

– flows match IP, generics match abstracted IP

Techniques extend to arbitrary sized rules

Instead,1. Focus on pair-wise rules (simpler is likelier)2. Group similar rules

– Eliminate weak rules between strongly connected groups– Transitive closure to read off clusters

Rule Mining

Mining for Rules

YX ⇒ YXXX n ⇒∧∧∧ K21O(f2) O(fn+1)

Rule ScoreRecursive Spectral Partitioning (VKV’00)

Digests 105—106 flows into 102—103 rule clustersDigests 105—106 flows into 102—103 rule clusters

…flowi.new flowj.present

...

PacketTrace

flow1 … flowK

time1present |new

…

timeR

Activity Matrix Rules

Recap: eXpose Mines for Rules

Learn all significant rules without prior knowledgeo Scoring function for rule significanceo Avoids observer bias, yet stays feasible by focusing on high hit-rateo Algorithms to mine and prune

Rule Clusters

Contributions

Related Work

Semi-Automated Discovery of App. Session Structure (KJPK’06)Sherlock (Diagnosing Performance Problems, BCGKMZ’07)Autofocus (ESV’03)BLINC (KPF’05)Stepping Stones (ZP’00)

Learn all significant rules without prior knowledgeo Avoids observer bias, yet stays feasible by focusing on high hit-rateo Scoring function for rule significanceo Algorithms to mine and prune

Results

Evaluation Setup

• Traces at access and internal server-facing links– Packet Headers, Connection Records (Bro), some anon.

• Operational n/w with 103 clients, diverse traffic mix• Corroborated on test-bed traffic & vetted by admins.• Ran eXpose on a 2.4GHz x86 with 8GB RAM

Inside MicrosoftBefore CSAIL’s ServersAccess Link of Conf. LANsCSAIL’s Access

• Dependencies for Major Applications

Rules Discovered by eXpose

Client.* – Mail.135

Client.* – DC.88 Client.* – Mail.X

Client.* – PFS1.X Client.* – PFS2.X Client.* – Proxy.80

email @ microsoftemail @ microsoft

Rules Discovered by eXpose• Dependencies for Major Applications

afs @ csailafs @ csail

C.7001 – Root.7003

C.7001 – *.*

C.7001 – AFS1.7000

C.7001 – AFS2.7000 AFS1.7000 – Root.7002

Rules Discovered by eXpose• Dependencies for Major Applications

– web, e-mail, file-servers, IM, print, video broadcast

web @ microsoftweb @ microsoft

Proxy1.80 – *.*

Proxy2.80 – *.* Proxy3.80 – *.*

Proxy4.80 – *.*

Rules Discovered by eXpose• Dependencies for Major Applications

– web, e-mail, file-servers, IM, print, video broadcast• Configuration Errors & Other Badness

Client.* – MailServer.25

Client.113 – MailServer.*

smtp + IDENT @ csailsmtp + IDENT @ csail

• Dependencies for Major Applications– web, e-mail, file-servers, IM, print, video broadcast

• Configuration Errors & Other Badness– IDENT, Legacy emails, ssh scans, wingate

Rules Discovered by eXpose

Legacy email ids @ csailLegacy email ids @ csail

UnivMail.* – Old2.25

UnivMail.* – Old1.25

UnivMail.* – Old3.25

Rules Discovered by eXpose• Dependencies for Major Applications

– web, e-mail, file-servers, IM, print, video broadcast• Configuration Errors & Other Badness

– IDENT, Legacy emails, ssh scans, wingate• Rules for stuff we didn’t know before

Nagios monitors @ csailNagios monitors @ csail

Nagios.7001 – AFS1.7000

Nagios.7001 – AFS2.7000

Nagios.* – Mail2.25

Nagios.* – Mail1.25

Rules Discovered by eXpose• Dependencies for Major Applications

– web, e-mail, file-servers, IM, print, video broadcast• Configuration Errors & Other Badness

– IDENT, Legacy emails, ssh scans, wingate• Rules for stuff we didn’t know before

– Nagios, LLMNR, iTunesLink level multicast name resolution @ hotspotsLink level multicast name resolution @ hotspots

H.* – DNS.53

H.137 – Wins.137

H.* – Multicast.5355Black box: Little prior knowledge about servers, applications, or users Can evolve

Black box: Little prior knowledge about servers, applications, or users Can evolve

Correctness & Completeness• False Positives

– 13% of rule-clusters in CSAIL trace, we couldn’t explain• False Negatives

– Main CSAIL Web Server (too many different activities)– Dependencies on Personal Web Pages (too few traffic)– PlanetLab Traffic (punted)

• Other Limitations– IPSec, Anonymized, Cover Traffic

• Extensions– Rules repeat over time, and across traces– Application whitelisting, Customize Generics

Time to Mine for Rules

At CSAIL’s access link, high fan-out with many distinct flows

Stream Mining Appears Feasible!Stream Mining Appears Feasible!

# Flows (x 106)

.6

.2

.6

.9

2.8

Packet TraceRules for frequently reoccurring flow sets

Learn all significant rules with no specific knowledgeo Avoids observer bias, but feasible by focusing on high hit-rateo Scoring function for rule significanceo Algorithms to mine and prune

Empirical validation on enterprise traces• found configurations & protocols that we didn’t know existed• learnt rules for actual behavior of applications• found config. errors, bot scans, infected machines

eXpose

http://research.microsoft.com/~srikanth

Backup

Rule Score (Modified JMeasure)

# of

Dis

cove

red

Rule

sExpanding Search Space (# of flows)…

… exposes few significant rules!… exposes few significant rules!

Expanding Search Space (# of flows)…

# Top Active Flows # Top Active Flows

Tim

e to

Min

e Ru

les

(s)

Mem

ory

Foot

prin

t (m

illio

n ru

les)

… exposes few rules & costs a lot in time, memory… exposes few rules & costs a lot in time, memory

Varying Size of Time Windows#

of D

isco

vere

d Ru

les

Rule Score (Modified JMeasure)

All window sizes in [.25s, 2s] produce similar rules!All window sizes in [.25s, 2s] produce similar rules!

For all rules X Y

Prob. (X)Prob. (Y)

Join

t Pro

babi

lity