01/04/2007ecs236 winter 20071 Intrusion Detection ecs236 Winter 2007: Intrusion Detection #2: Anomaly Detection Dr. S. Felix Wu Computer Science Department.

01/04/2007 ecs236 winter 2007 1

ecs236 Winter 2007:

Intrusion DetectionIntrusion Detection#2: Anomaly Detection

Dr. S. Felix Wu

Computer Science Department

University of California, Davishttp://www.cs.ucdavis.edu/~wu/

[email protected]

01/04/2007 ecs236 winter 2007 2

Intrusion DetectionIntrusion Detection

IntrusionDetection

Model

Input eventsequence Results

Pattern matching

01/04/2007 ecs236 winter 2007 3

Scalability of DetectionScalability of Detection

Number of signatures, amount of analysis Unknown exploits/vulnerabilities

01/04/2007 ecs236 winter 2007 4

Anomaly vs. SignatureAnomaly vs. Signature Signature Intrusion (Bad things happen!!)

– Misuse produces observable bad effect– Specify and look for bad behaviors

Anomaly Intrusion (Good things did not happen!!)– We know what our normal behavior is– Looking for an deviation from the normal

behavior, raise early warning

01/04/2007 ecs236 winter 2007 5

Reasons for “AND”Reasons for “AND”

Unknown attacks (insider threat) Better scalability

– AND target/vulnerabilities– SD exploits

01/04/2007 ecs236 winter 2007 6

Another definition…Another definition… Signature-based detection

– Predefine the signatures of anomalies– Pattern matching

Statistics-based detection– Build statistics profile for expected behaviors

– Compare testing behaviors with expected behaviors

– Significant deviation

Convert our limited/partial understanding/modeling about the target system or protocol into detection heuristics (i.e., BUTTERCUP signatures)

Based on our experience, select a set of “features” that will likely to distinguish expected from unexpected behavior.

01/04/2007 ecs236 winter 2007 7

What is “vulnerability”?What is “vulnerability”?

01/04/2007 ecs236 winter 2007 8

What is “vulnerability”?What is “vulnerability”?

Signature Detectioncreate “effective/strong/scaleable” signatures

Anomaly Detectiondetect/discover “unknown vulnerabilities”

01/04/2007 ecs236 winter 2007 9

ANDAND(ANomaly Detection)(ANomaly Detection)

Unknown Vulnerabilities/Exploits Insider Attacks

Understand How and Why these things happened

Understand the limit of AND from both sides

01/04/2007 ecs236 winter 2007 10

What is an anomaly?What is an anomaly?

01/04/2007 ecs236 winter 2007 11

Intrusion DetectionIntrusion Detection

IntrusionDetection

Model

Input eventsequence Results

Pattern matching

01/04/2007 ecs236 winter 2007 12

Anomaly DetectionAnomaly Detection

IntrusionDetection

Input eventsequence

Pattern matching

01/04/2007 ecs236 winter 2007 13

For each sample of the statistic measure, X

(0, 1]

40%

(1, 3]

30%

(3, 15]

20%

(15, +)

10%

Input Events

iq 1+iq

SAND

01/04/2007 ecs236 winter 2007 14

quantify the anomalies

alarm generation

threshold control

raw events

long term profile

0 5 10 15 20 25 300

“But, which feature(s) to profile??”

function F

01/04/2007 ecs236 winter 2007 15

Statistic-based ANomaly DetectionStatistic-based ANomaly Detection(SAND)(SAND)

choose a parameter (a random variable hopefully without any assumption about its probabilistic distribution)

record its statistical “long-term” profile check how much, quantitatively, its short-term

behavior deviates from its long term profile set the right threshold on the deviation to raise

alarms

01/04/2007 ecs236 winter 2007 16

decay

update

clean

compute thedeviation

alarm generation

threshold control

timer control

raw events long term profile

0 5 10 15 20 25 300

01/04/2007 ecs236 winter 2007 17

False Positive & NegativeFalse Positive & Negative

Long term profile Quantitative measure of the deviation

between long term and target of detection Threshold-based control

01/04/2007 ecs236 winter 2007 18

Long-term ProfileLong-term Profile

Category, C-Training learn the aggregate distribution of a statistic

measure

Q Statistics, Q-Training

learn how much deviation is considered normal

Threshold

01/04/2007 ecs236 winter 2007 19

Long-term Profile: C-TrainingLong-term Profile: C-TrainingFor each sample of the statistic measure, X

(0, 50]

20%

(50, 75]

30%

(75, 90]

40%

(90, +)

10% k bins Expected Distribution, P1 P2 ... Pk , where

Training time: months ∑= =ki ip1 1

01/04/2007 ecs236 winter 2007 20

Long-term Profile: Q-Training (1)Long-term Profile: Q-Training (1)

For each sample of the statistic measure, X

(0, 50]

20%

(50, 75]

40%

(75, 90]

20%

(90, +)

20%

k bins, samples fall into bin samples in total ( ) Weighted Sum Scheme with the fading factor s

iY ′ thiN ′ ∑= ′=′k

i i NY1

01/04/2007 ecs236 winter 2007 21

ThresholdThreshold Predefined threshold, If Prob(Q>q) < , raise alarm

0

0.08

0 5 10 15 20 25 30

Q bins

Probability TH_redTH_yellow

∑= ×′

×′−′=

k

i i

ii

pN

pNYQ

1

2)(

01/04/2007 ecs236 winter 2007 22


Deviation:

Example:

Qmax

the largest value among all Q values

∑= ×′

×′−′=

k

i i

ii

pN

pNYQ

1

2)(

33.21.010

)1.0102(

4.010

)4.0102(

3.010

)3.0104(

2.010

)2.0102( 2222

=××−

+××−

+××−

+××−

=Q

01/04/2007 ecs236 winter 2007 23


Q Distribution [0, Qmax) is equally divided into 31 bins and

the last bin is [Qmax, +)distribute all Q values into the 32 bins

01/04/2007 ecs236 winter 2007 24

Q-MeasureQ-Measure

Deviation:

Example:

Qmax

the largest value among all Q values

∑= ×′

×′−′=

k

i i

ii

pN

pNYQ

1

2)(

33.21.010

)1.0102(

4.010

)4.0102(

3.010

)3.0104(

2.010

)2.0102( 2222

=××−

+××−

+××−

+××−

=Q

01/04/2007 ecs236 winter 2007 25

iq 1+iq

01/04/2007 ecs236 winter 2007 26

ThresholdThreshold Predefined threshold, If Prob(Q>q) < , raise alarm

0

0.08

0 5 10 15 20 25 30

Q bins

Probability TH_redTH_yellow

∑= ×′

×′−′=

k

i i

ii

pN

pNYQ

1

2)(

Falsepositive

01/04/2007 ecs236 winter 2007 27

Heidelberg

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

NCU

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

SingNet

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

UIUC

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

01/04/2007 ecs236 winter 2007 28

MathematicsMathematics Many other techniques:

– Training/learning– detection

01/04/2007 ecs236 winter 2007 29

Statistical Profiling Long-Term profile:Long-Term profile:

capture long-term behavior of a particular capture long-term behavior of a particular statistic measurestatistic measure

e.g., update once per daye.g., update once per day half-life: 30 updateshalf-life: 30 updates

recent 30: recent 30: 50%50% 31-60: 31-60: 25%25% the newer contributes morethe newer contributes more

01/04/2007 ecs236 winter 2007 30

Statistical Pros and ConsStatistical Pros and Cons Slower to detect - averaging window Very good for unknown attacks - as long as

“relevant measures” are chosen Environment (protocol, user, etc)

dependency– Need good choices on statistical measures– Statistical profiles might be hard to build– Thresholds might be hard to set

01/04/2007 ecs236 winter 2007 31

decay

update

clean


alarm generation

threshold control

timer control


0 5 10 15 20 25 300

01/04/2007 ecs236 winter 2007 32

Weighted Sum SchemeWeighted Sum Scheme Problems of Sliding Window Scheme

Keep the most recent N pieces of audit records

required resource and computing time are O(N)

12

,2

12

1 +×==

≠×=

+×=

−=

−

−

∑

NYN

ijYY

YY

ki i

jj

ii

Assume K: number of bins Yi: count of audit records falls

into ith bin N: total number of audit records : fading factor

When Ei occurs, update

01/04/2007 ecs236 winter 2007 33

FTP Severs and ClientsFTP Severs and Clients

FTP Client

SHANG

FTP Servers

Heidelberg

NCU

SingNet

UIUC

01/04/2007 ecs236 winter 2007 34

Dropper AttacksDropper Attacks

P%P% Per (K,I,S)Ret (K,S)Ran (K)

Intentional or Unintentional??

01/04/2007 ecs236 winter 2007 35

Periodical Packet DroppingPeriodical Packet Dropping

Parameters (K, I, S) K, the total number of dropped packets in a connection I, the interval between two consecutive dropped packets S, the position of the first dropped packet.

Example (5, 10, 4) 5 packets dropped in total 1 every 10 packets start from the 4th packet The 4th, 14th, 24th, 34th and 44th packet will be dropped

01/04/2007 ecs236 winter 2007 36

Retransmission Packet DroppingRetransmission Packet Dropping

Parameters (K, S) K, the times of dropping the packet's retransmissions S, the position of the dropped packet

Example (5, 10) first, drops the 10th packet then, drops the retransmissions of the 10th packet 5

times

01/04/2007 ecs236 winter 2007 37

Random Packet DroppingRandom Packet Dropping

Parameters (K) K, the total number of packets to be dropped in a

connection

Example (5) randomly drops 5 packets in a connection

01/04/2007 ecs236 winter 2007 38

Experiment SettingExperiment Setting

FTP

Internet

Divert Socket

FTP Client

xyz.zip 5.5M

FTP Server

Attack Agent

Data Packets

01/04/2007 ecs236 winter 2007 39

Impacts of Packet Dropping On Impacts of Packet Dropping On Session DelaySession Delay

5663.4 66

218.4

98.6

108.2

125.8

250.9

62.6

77.1

86.9

260.3

23.626.5

44.6

183.9

0

50

100

150

200

250

300

Session Delay (s)

Heidelberg NCU SingNet UIUC

Normal

RanPD(7)

PerPD(7, 4, 5)

RetPD(7, 5)

01/04/2007 ecs236 winter 2007 40

Compare Impacts of Dropping Compare Impacts of Dropping PatternsPatterns

0

50

0

-10 40

Num ber

Session

He ide lbe rg

0

50

100

150

200

250

300

350

400

450

500

0 10 20 30 40

Nu m b er o f victim p ack ets

Session delay

PerPD

RanPD

RetPD

NCU

0

50

100

150

200

250

300

350

400

450

500

0 10 20 30 40


Session delay

PerPD

RanPD

RetPD

SingNe t

0

50

100

150

200

250

300

350

400

450

500

0 10 20 30 40


Session delay

PerPD

RanPD

RetPD

UIUC

0

50

100

150

200

250

300

350

400

450

500

0 10 20 30 40


Session delay

PerPD

RanPD

RetPD

PerPD: I=4, S=5

RetPD: S=5

01/04/2007 ecs236 winter 2007 41

bone

fire

redwing

light

152.1.75.0

192.168.1.0

172.16.0.0

UDP flood

FTP data

TFN agents

TFN target

FTP client

FTP server

congestion

air

TFN master

01/04/2007 ecs236 winter 2007 42

flood 1, Stop 20

0

2

4

6

8

10

12

0 20 40 60 80 100

Time (s)

Number of Lost Packets

flood 1, Stop 5

0

2

4

6

8

10

12

0 20 40 60 80 100

Time (s)


flood 5, Stop 10

0

2

4

6

8

10

12

0 20 40 60 80 100

Time (s)


flood 5, Stop 2

0

2

4

6

8

10

12

0 20 40 60 80 100

Time (s)


01/04/2007 ecs236 winter 2007 43

TDSAM Experiment SettingTDSAM Experiment Setting

FTP

Internet

Divert Socket

FTP Client

xyz.zip 5.5M

FTP Server

Attack Agent

TDSAM

Data Packets

p1, p2, p3, p5, p4max

reordering counting

01/04/2007 ecs236 winter 2007 44

Heidelberg

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

NCU

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

SingNet

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

UIUC

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0 5 10 15 20 25 30 35Q bins

Probability

01/04/2007 ecs236 winter 2007 45

He ide lbe rg

0

0.05

0.1

0.15

0.2

0.25

0.3

0 5 10 15 20 25 30 35

Q bins

Probability

NCU

0

0.05

0.1

0.15

0.2

0.25

0.3

0 5 10 15 20 25 30 35

Q bins

Probability

SingNet

0

0.05

0.1

0.15

0.2

0.25

0.3

0 5 10 15 20 25 30 35

Q bins

Probability

UIUC

0

0.05

0.1

0.15

0.2

0.25

0.3

0 5 10 15 20 25 30 35

Q bins

Probability

01/04/2007 ecs236 winter 2007 46

Results: Position MeasureResults: Position MeasureHeidelberg NCU SingNet UIUCPosition

nbin=5 DR MR DR MR DR MR DR MR

Normal* - 4.0% - 5.4% - 3.5% - 6.5% -

(10, 4, 5) 99.7% 0.3% 100% 0% 100% 0.0% 100% 0%

(20, 4, 5) 100% 0% 98.1% 1.9% 99.2% 0.8% 100% 0%

(40, 4, 5) 96.6% 3.4% 100% 0% 100% 0% 98.5% 1.5%

(20, 20, 5) 100% 0% 100% 0% 100% 0 % 100% 0%

(20, 100, 5) 98.9% 1.1%. 99.2% 0.8% 99.6% 0.4% 99.1% 0.9%

(20, 200, 5) 0% 100% 76.5% 23.5% 1.5% 98.5% 98.3% 1.7%

PerPD

(100, 40, 5) 0.2% 99.8% 0% 100% 0% 100% 100% 0%

RetPD (5, 5) 84.9% 15.1% 81.1% 18.9% 94.3% 5.7% 97.4% 2.6%

10 0% 100% 42.3% 57.7% 0% 100% 0% 100%RanPD

40 0% 100% 0% 100% 0% 100% 0% 100%

5 98.6% 1.4% 100% 0% 98.2% 1.8% 100% 0%Intermittent

(10, 4, 5) 50 34.1% 65.9% 11.8% 88.2% 89.4% 10.6% 94.9% 5.1%

01/04/2007 ecs236 winter 2007 47

Results: Delay MeasureResults: Delay MeasureHeidelberg NCU SingNet UIUCDelay


Normal* - 1.6% - 7.5% - 2.1% - 7.9% -

(10, 4, 5) 97.4% 2.6% 95.2% 4.8% 94.5% 5.5% 99.2% 0.8%

(20, 4, 5) 99.2% 0.8% 98.5% 1.5% 100% 0% 100% 0%

(40, 4, 5) 100% 0% 100% 0% 100% 0% 100% 0%

(20, 20, 5) 96.3% 3.7% 100% 0% 92.6% 7.4% 98.9% 1.1%

(20, 100, 5) 100% 0% 95.3% 4.7% 98.7% 1.3% 100% 0%

(20, 200, 5) 98.6% 1.4% 99% 1% 97.1% 2.9% 100% 0%

PerPD

(100, 40, 5) 100% 0% 100% 0% 100% 0% 100% 0%

RetPD (5, 5) 100% 0% 100% 0% 100% 0% 100% 0%

10 74.5% 25.5% 26.8% 73.2% 67.9% 32.1% 99.5% 0.5%RanPD

40 100% 0% 100% 0% 100% 0% 100% 0%

5 25.6% 74.4% 0% 100% 0% 100% 97.3% 2.7%Intermittent

(10, 4, 5) 50 0% 100% 24.9% 75.1% 0% 100% 3.7% 96.3%

01/04/2007 ecs236 winter 2007 48

Results: NPR MeasureResults: NPR MeasureHeidelberg NCU SingNet UIUCNPR


Normal* - 4.5% - 5.8% - 8.2% - 2.9% -

(10, 4, 5) 0% 100% 14.4% 85.6% 29.1% 70.9% 100% 0%

(20, 4, 5) 83.1% 16.9% 94.2% 5.8% 95.2% 4.8% 100% 0%

(40, 4, 5) 100% 0% 97.4% 2.6% 100% 0% 100% 0%

(20, 20, 5) 91.6% 8.4% 92% 8% 93.5% 6.5% 100% 0%

(20, 100, 5) 94.3% 5.7% 92.2% 7.8% 96.4% 3.6% 100% 0%

(20, 200, 5) 0% 100% 96.5% 3.5% 94.8% 5.2% 100% 0%

PerPD

(100, 40, 5) 100% 0% 100% 0% 100% 0% 100% 0%

RetPD (5, 5) 0% 100% 84.7% 15.3% 23.9% 76.1% 46.5% 53.5%

10 0% 100% 0% 100% 100% 0% 100% 0%RanPD

40 100% 0% 100% 0% 100% 0% 100% 0%

5 0% 100% 0% 100% 82.2% 17.8% 100% 0%Intermittent

(10, 4, 5) 50 0% 100% 1% 99% 40% 60% 64.8% 35.2%

01/04/2007 ecs236 winter 2007 49

Results (good and bad)Results (good and bad)

False Alarm Rate less than 10% in most cases, the highest is 17.4%

Detection Rate Position: good on RetPD and most of PerPD

at NCU, 98.7% for PerPD(20,4,5), but 0% for PerPD(100, 40, 5) in which dropped packets are evenly distributed

Delay: good on those significantly change session delay, e.g., RetPD, PerPD with a large value of K

at SingNet, 100% for RetPD(5,5), but 67.9% for RanPD(10)

NPR: good on those dropping many packets at Heidelberg, 0% for RanPD(10), but 100% for RanPD(40)

01/04/2007 ecs236 winter 2007 50

Performance AnalysisPerformance Analysis

Good sites correspond to a high detection rate. stable and small session delay or packet reordering

e.g., using Delay Measure for RanPD(10): UIUC (99.5%) >

Heidelberg(74.5%) > SingNet (67.9%) > NCU (26.8%)

How to choose the value of nbin is site-specific e.g., using Position Measure, lowest false alarm rate occurs when

nbin= 5 at Heidelberg(4.0%) and NCU(5.4%), 10 at UIUC(4.5%) and

20 at SingNet(1.6%)

01/04/2007 ecs236 winter 2007 51

decay

update

clean


alarm generation

threshold control

timer control


0 5 10 15 20 25 300

01/04/2007 ecs236 winter 2007 52

decay

update

clean

cognitivelyidentify thedeviation

alarm identification

InformationVisualizationToolkit

raw events cognitive profile

01/04/2007 ecs236 winter 2007 53


01/04/2007 ecs236 winter 2007 54

What is an anomaly?What is an anomaly? The observation of a target system is

inconsistent, somewhat, with the expected conceptual model of the same system

01/04/2007 ecs236 winter 2007 55

What is an anomaly?What is an anomaly? The observation of a target system is

inconsistent, somewhat, with the expected conceptual model of the same system

And, this conceptual model can be ANYTHING.– Statistical, logical, or something else

01/04/2007 ecs236 winter 2007 56

Model vs. ObservationModel vs. Observationthe Model Anomaly Detection

Conflicts Anomalies

It could be an attack, but it might well be misunderstanding!!

01/04/2007 ecs236 winter 2007 57

The ChallengeThe Challenge

Events

Expected Behavior Model

Anomaly Detection

Knowledge about the Target

False Positives & Negatives

01/04/2007 ecs236 winter 2007 58


Events


Anomaly Detection

01/04/2007 ecs236 winter 2007 59


Events


Anomaly Detection

Knowledge about the Target

01/04/2007 ecs236 winter 2007 60

Model vs. ObservationModel vs. Observationthe Model Anomaly Detection

Conflicts Anomalies

It could be an attack, but it might well be misunderstanding!!

01/04/2007 ecs236 winter 2007 61

ChallengeChallenge We know that the detected anomalies can

be either true-positive or false-positive. We try all our best to resolve the puzzle by

examining all information available to us. But, the “ground truth” of these anomalies

is very hard to obtain– even with human intelligence

01/04/2007 ecs236 winter 2007 62

Problems with ANDProblems with AND We are not sure about whatever we want to

detect… We are not sure either when something is

caught… We are still in the dark… at least in many

cases…

01/04/2007 ecs236 winter 2007 63

Anomaly ExplanationAnomaly Explanation How will a human resolve the conflict?

The Power of Reasoning and Explanation– We detected something we really want to detect

reducing false negative– Our model can be improved reduce false

positive

01/04/2007 ecs236 winter 2007 64

Without ExplanationWithout Explanation AND is not as useful?? Knowledge is the power to utilize

information!– Unknown vulnerabilities– Root cause analysis– Event correlation

01/04/2007 ecs236 winter 2007 65

Anomaly ExplanationAnomaly Explanation

the Model Anomaly Detection

Anomaly Analysis and Explanation

EBL

Explaining both the attack and the normal behavior

01/04/2007 ecs236 winter 2007 66

ExplanationExplanation

SimulationExperiments

OrObservatinon

Conflicts Anomalies

01/04/2007 ecs236 winter 2007 67

the Modelmodel-based

event analysis

observed system events

SBL-basedAnomalyDetection

analysisreports

ExampleSelection

Explanation Based

Learning

modelupdate

01/04/2007 ecs236 winter 2007 68

AND AND EXPAND EXPAND

Anomaly Detection– Detect– Analysis and Explanation– Application

01/04/2007 ecs236 winter 2007 69

01/04/2007 ecs236 winter 2007 70

01/04/2007 ecs236 winter 2007 71

01/04/2007 ecs236 winter 2007 72

01/04/2007ecs236 winter 20071 Intrusion Detection ecs236 Winter 2007: Intrusion Detection #2: Anomaly Detection Dr. S. Felix Wu Computer Science Department.

Documents

01042007ecs236 winter

signature detection

scalability of detection

detection heuristics

signaturebased detection

input events sand slide

early warning slide

function f slide