1

Building A Trustworthy, Secure, And Privacy Preserving Network

Bharat BhargavaCERIAS Security CenterCWSA Wireless Center

Department of CS and ECEPurdue University

Supported by NSF IIS 0209059, NSF IIS 0242840 ,

NSF ANI 0219110, CISCO, Motorola, IBM

2

Research Team

• Faculty Collaborators– Dongyan Xu, middleware and privacy– Mike Zoltowski, smart antennas, wireless security– Sonia Fahmy, Internet security – Ninghui Li, trust– Cristina Nita-Rotaru, Internet security

• Postdoc– Lezsek Lilien, privacy and vulneribility– Xiaoxin Wu, wireless security– Jun Wen, QoS– Mamata Jenamani, privacy

• Ph.D. students– Ahsan Habib, Internet Security– Mohamed Hefeeda, peer-to-peer– Yi Lu, wireless security and congestion control– Yuhui Zhong, trust management and fraud– Weichao Wang, security of ad hoc networks

More information at http://www.cs.purdue.edu/people/bb

3

Motivation

• Lack of trust, privacy, security, and reliability impedes information sharing among distributed entities.– San Diego supercomputer center detected 13,000

DoS attacks in a three-week period [eWeek, 2003]– Internet attacks in February, 2004 caused an

estimated $68 billion to $83 billion in damages worldwide [British Computer Security Report]

– Business losses due to privacy violations• Online consumers worry about revealing personal data• This fear held back $15 billion in online revenue in 2001

– 52,658 reported system crashes caused by software vulnerabilities in 2002 [Express Computers 2002]

4

Research is required for the creation of knowledge and learning in secure networking, systems, and applications.

5

• Enable the deployment of security sensitive applications in the pervasive computing and communication environments.

Goal

6

Problem Statement

• A trustworthy, secure, and privacy preserving network platform must be established for trusted collaboration. The fundamental research problems include:– Trust management– Privacy preserved interactions– Dealing with a variety of attacks and frauds in

networks– Intruder identification in ad hoc networks (focus

of this seminar)

7

Applications/Broad Impacts

• Guidelines for the design and deployment of security sensitive applications in the next generation networks– Data sharing for medical research and treatment– Collaboration among government agencies for

homeland security– Transportation system (security check during travel,

hazardous material disposal)– Collaboration among government officials, law

enforcement and security personnel, and health care facilities during bio-terrorism and other emergencies

8

A. Trust Formalization• Problem

– Dynamically establish and update trust among entities in an open environment.

• Research directions– Handling uncertain evidence– Modeling dynamic trust– Formalization and detection of fraud

• Challenges– Uncertain information complicates the inference procedure.– Subjectivity leads to various interpretations toward the same

information.– The multi-faceted and context-dependent characteristics of trust

require tradeoff between representation comprehensiveness and computation simplicity of the trust model.

9

Trust Enhanced Role Assignment (TERA) Prototype

• Trust enhanced role mapping (TERM) server assigns roles to users based on – Uncertain & subjective evidence

– Dynamic trust

• Reputation server – Dynamic trust information repository– Evaluate reputation from trust information by using

algorithms specified by TERM server

Prototype and demo are available at

http://www.cs.purdue.edu/homes/bb/NSFtrust/

10

TERA Architecture

T E R M s er v er

T E R M s er v er

T r u s t b as ed o n b eh av io r s

T r u s t b as ed o n b eh av io r s

R ep u ta tio n

R ep u ta tio n

R ep u ta tio n s er v er

Alic e

Bo b

T E R A

R o le r eq u es t

As s ig n ed r o le

R o le r eq u es t

As s ig n ed r o le

R BAC en h an c edap p lic a tio n s er v er

R BAC en h an c edap p lic a tio n s er v er

Us er 's b eh av io r

Us er 's b eh av io r

I n te r ac tio n s

I n te r ac tio n s

11

Trust Enhanced Role Mapping (TERM) Server

• Evidence rewriting

• Role assignment– Policy parser – Request processor & inference engine– Constraint enforcement

• Policy base

• Trust information management– User behavior modeling – Trust production

12

TERM Server

TERM

Credential Manager

Assign role

Credentials provided / retrieved

Role Assignment

Evidence statement

Evidence statement

Evidence Rewriting Trust toward

issuer

Trust toward user/issuer

Trust

Information Management

Behaviors

Policy Base

Role-assignment Policy

Role-assignment policies

Reputation

user

Reputation server

Policy maker

Application server

Trust information

Request role

13

Fraud Formalization and Detection

• Model fraud intention– Uncovered deceiving intention– Trapping intention– Illusive intention

• Fraud detection– Profile-based anomaly detection

• Monitor suspicious actions based upon the established patterns of an entity

– State transition analysis• Build an automaton to identify activities that lead

towards a fraudulent state

14

Model Fraud Intentions

• Uncovered deceiving intention– Satisfaction ratings

are stably low. – Ratings vary in a

small range over time.

15

Model Fraud Intentions

• Trapping intention– Rating sequence can

be divided into two phases: preparing and trapping.

– A swindler behaves well to achieve a trustworthy image before he conducts frauds.

16

Model Fraud Intentions

• Illusive intention– A smart swindler

attempts to cover the bad effects by intentionally doing something good after misbehaviors.

– Process of preparing and trapping is repeated.

17

B. Private and Trusted Interactions

• Problem– Preserve privacy, gain trust, and control dissemination of data

• Research directions– Dissemination of private data– Privacy and trust tradeoff– Privacy metrics

• Challenges– Specify policies through metadata and establish guards as

procedures– Efficient implementation– Estimate privacy depending on who will get this information,

possible uses of this information, and information disclosed in the past

– Privacy metrics are usually ad hoc and customized

Detail slides at http://www.cs.purdue.edu/homes/bb/priv_trust_cerias.ppt

18

Preserving Privacy in Data Dissemination

• Design self-descriptive private objects

• Construct a mechanism for apoptosis of private objects

apoptosis = clean self-destruction

• Develop proximity-based evaporation of private objects

• Develop schemes for data distortions

19

Privacy Metrics

• Determine the degree of data privacy– Size of anonymity set metrics– Entropy-based metrics

• Privacy metrics should account for:– Dynamics of legitimate users– Dynamics of violators– Associated costs

20

Size of Anonymity Set Metrics

• The larger set of indistinguishable entities, the lower probability of identifying any one of them– Can use to ”anonymize” a selected private attribute value

within the domain of its all possible values

“Hiding in a crowd”

“More” anonymous (1/n)

“Less” anonymous (1/4)

21

Dynamics of Entropy

• Decrease of system entropy with attribute disclosures (capturing dynamics)

– When entropy reaches a threshold (b), data evaporation can be invoked to increase entropy by controlled data distortions

– When entropy drops to a very low level (c), apoptosis can be triggered to destroy private data

– Entropy increases (d) if the set of attributes grows or the disclosed attributes become less valuable – e.g., obsolete or more data now available

(a)

(b)

(c) (d)

Disclosed attributes

H*

Allattribut

es

Entropy

Level

22

Private and Trusted System (PRETTY) Prototype

(1)

[2a]

(3) User Role

[2b] [2d][2c1]

[2c2]

(2)

(4)

TERA = Trust-Enhanced Role Assignment

23

C. Tomography Research• Problem

– Defend against denial of service attacks – Optimize the selection of data providers in peer-to-

peer systems

• Research Directions– Stripe based probing to infer individual link loss by

edge-to-edge measurements– Overlay based monitoring to identify congested links

by end-to-end path measurement– Topology inference to estimate available bandwidth

by path segment measurements

24

Defeating DoS Attacks in Internet

25

Overlay-based Monitoring

• Do not need individual link loss to identify all congested links

• Edge routers form an overlay network for probing. Each edge router probe part of the network

• Problem statement– Given topology of a network domain, identify

which links are congested and possibly under attack

26

Attack Scenarios

(a) Changing delay pattern due to attack

(b) Changing in loss pattern due to attack

Time (sec) Time (sec)

D

ela

y (

ms)

L

oss

Rati

o

27

Identified Congested Links

(a) Counter clockwise probing (b) Clockwise probing

Probe46 in graph (a) and Probe76 in graph (b) observe high losses, which means link C4 E6 is congested.

Time (sec) Time (sec)

Loss

Rati

o

L

oss

Rati

o

28

Probing: Simple Method

(a) Topology (b) Overlay (c) internal links

Congested link

29

D. Intruder Identification in Adhoc On-demand Distance Vector (AODV)

• Problem – AODV are vulnerable to various attacks such as false

distance vector, false destination sequence, and wormhole attacks

– Detecting attacks without identifying and isolating the malicious hosts leaves the security mechanisms in a passive mode

• Challenges– Locate the sources of attacks in a self-organized

infrastructure– Combine local decisions with knowledge from other

hosts to achieve consistent conclusions on the malicious hosts

30

• Related Work– Vulnerability model of ad hoc routing protocols [Yang

et al., SASN ’03]– A generic multi layer integrated IDS structure [Zhang

and Lee, MobiCom ’00]– IDS combining with trust [Albert et al., ICEIS ’02]– Information theoretic measures using entropy

[Okazaki et al., SAINT ’02]– SAODV adopts both hash chain and digital signature

to protect routing information [Zapata et al, WiSe’03]– Security-aware ad hoc routing [Kravets et al,

MobiHOC’01]

31

Ideas

• Monitor the sequence numbers in the route request packets to detect abnormal conditions

• Apply reverse labeling restriction to identify and isolate attackers

• Combine local decisions with knowledge from other hosts to achieve consistent conclusions

• Combine with trust assessment methods to improve robustness

32

Introduction to AODV

• Introduced in 97 by Perkins at NOKIA, Royer at UCSB

• 12 versions of IETF draft in 4 years, 4 academic implementations, 2 simulations

• Combines on-demand and distance vector• Broadcast Route Query, Unicast Route Reply• Quick adaptation to dynamic link condition

and scalability to large scale network• Support multicast

33

Route Discovery in AODV (An Example)

S

D

S1

S2

S3

S4

Route to the source

Route to the destination

34

Attacks on AODV• Route request flooding

– query non-existing host (RREQ will flood throughout the network)

• False distance vector– reply “one hop to destination” to every request and select a

large enough sequence number

• False destination sequence number– select a large number (even beat the reply from the real

destination)

• Wormhole attacks– tunnel route request through wormhole and attract the data

traffic to the wormhole

• Coordinated attacks– The malicious hosts establish trust to frame other hosts, or

conduct attacks alternatively to avoid being identified

35

False Destination Sequence Attack

D

S S1

S2 M

S3

RREQ(D, 3)

RREQ(D, 3)

RREQ(D, 3)

RREQ(D, 3)

RREP(D, 5)

RREP(D, 20)

Packets from S to D are sinking at M. Node movement breaks the path from S to M (trigger route rediscovery).

36

During Route Rediscovery, False Destination Sequence Attack Is Detected

D

S S1

S2 M

S3

S4

RREQ(D, 21)

(1). S broadcasts a request that carries the old sequence + 1 = 21

(2) D receives the RREQ. Local sequence is 5, but the sequence in RREQ is 21. D detects the false desti-nation sequence attack.

Propagation of RREQ

37

Reverse Labeling Restriction (RLR)

Blacklists are updated after an attack is detected.• Basic Ideas

• Every host maintains a blacklist to record suspicious hosts. Suspicious hosts can be released from the blacklist.

• The destination host will broadcast an INVALID packet with its signature when it finds that the system is under attack on sequence. The packet carries the host’s identification, current sequence, new sequence, and its own blacklist.

• Every host receiving this packet will examine its route entry to the destination host. If the sequence number is larger than the current sequence in INVALID packet, the presence of an attack is noted. The previous host that provides the false route will be added into this host’s blacklist.

38

D

S S1

S2M

S3

S4

BL {}

BL {S2}

BL {}BL {M}

BL {S1}

BL {}

D broadcasts INVALID packet with current sequence = 5, new sequence = 21. S3 examines its route table, the entry to D is not false. S3 forwards packet to S1. S1 finds that its route entry to D has sequence 20, which is > 5. It knows that the route is false. The hop which provides this false route to S1 was S2. S2 will be put into S1’s blacklist. S1 forwards packet to S2 and S. S2 adds M into its blacklist. S adds S1 into its blacklist. S forwards packet to S4. S4 does not change its blacklist since it is not involved in this route.

INVALID ( D, 5, 21, {}, SIGN )

Correct destination sequence number is broadcasted.

Blacklist at each host in the path is determined.

39

D4

D1

S3

S1

M

D3

S4

S2

D2

M attacks 4 routes (S1-D1, S2-D2, S3-D3, and S4-D4). When the first two false routes are detected, D3 and D4 add M into their blacklists. When later D3 and D4 become victim destinations, they will broadcast their blacklists, and every host will get two votes that M is malicious host.

[M]

[M]

[M]

[M]

Hosts closer to malicious site are in blacklists of multiple hosts. In the above figure, M is in four blacklists.

40

Combine Local Decisions with Knowledge from Other Hosts• When a host is destination of a route and is

victim by any malicious host, it will broadcast its blacklist.

• Each host obtains blacklists from victim hosts.

• If M is in multiple blacklists, M is classified as a malicious host based on certain threshold.

• Intruder is identified.• Trust values can be assigned to other hosts

based on past information.

41

D3

M1

S1

D1

Routing topology Reverse labeling procedure

Multiple attackers exist in the network. More routes are under attack. When the false routes are detected, more blacklists will be broadcasted.

D2

M2 M3

S2 S3

D3

M1

S1

D1D2

M2 M3

S2 S3

Acceleration in Intruder Identification

42

Reverse Labeling Restriction

• Update Blacklist by INVALID Packet• Next hop on the invalid route will be put into

local blacklist, a timer starts, and a counter increases. The time duration that the host stays in blacklist is exponential to the counter value.

• Labeling process will be conducted in the reverse direction of the false route.

• When timer expires, the suspicious host will be released from the blacklist and routing information from it will be accepted.

43

Deal With Hosts in Blacklist

• Packets from hosts in blacklist• Route request: If the request is from suspicious

hosts, ignore it. • Route reply: If the previous hop is suspicious and

the query destination is not the previous hop, the reply will be ignored.

• Route error: will be processed as usual. RERR will activate re-discovery, which will help to detect attacks on destination sequence.

• INVALID: if the sender is suspicious, the packet will be processed but the blacklist will be ignored.

44

Attacks of Malicious Hosts on RLR

• Attack 1: Malicious host M sends false INVALID packet• Because the INVALID packets are signed, it

cannot send the packets in other hosts’ name• If M sends INVALID in its own name

• If the reported sequence number is greater than the real sequence number, every host ignores this attack

• If the reported sequence number is less than the real sequence number, RLR will converge at the malicious host. M is included in blacklist of more hosts. M accelerated the intruder identification directing towards M.

45

• Attack 2: Malicious host M frames other innocent hosts by sending false blacklist• If the malicious host has been identified, the

blacklist will be ignored• If the malicious host has not been identified, this

operation can only make the threshold lower. If the threshold is selected properly, it will not impact the identification results.

• Combining trust can further limit the impact of this attack.

46

• Attack 3: Malicious host M only sends false destination sequence about some special host• The special host will detect the attack and

send INVALID packets.• Other hosts can establish new routes to the

destination by receiving the INVALID packets.

47

Experimental Studies of RLR

• The experiments are conducted using ns2.• Various network scenarios are formed by

varying the number of independent attackers, number of connections, and host mobility.

• The examined parameters include:– Packet delivery ratio– Identification accuracy: false positive and

false negative ratio– Communication and computation overhead

48

Simulation Parameter

Simulation duration 1000 seconds

Simulation area 1000 * 1000 m

Number of mobile hosts 30

Transmission range 250 m

Pause time between the host reaches current target and moves to next target

0 – 60 seconds

Maximum speed 5 m/s

Number of CBR connection 25/50

Packet rate 2 pkt / sec

49

Experiment 1: Measure the Changes in Packet Delivery Ratio

Purpose: investigate the impacts of host mobility, number of attackers, and number of connections on the performance improvement brought by RLR

Input parameters: host pause time, number of independent attackers, number of connections

Output parameters: packet delivery ratioObservation: When only one attacker exists in the

network, RLR brings a 30% increase in the packet delivery ratio. When multiple attacker exist in the system, the delivery ratio will not recover before all attackers are identified.

50

Increase in Packet Delivery Ratio: Single Attacker

X-axis is host pause time, which evaluates the mobility of host. Y-axis is delivery ratio. 25 connections and 50 connections are considered. RLR brings a 30% increase in delivery ratio. 100% delivery is difficult to achieve due to network partition, route discovery delay and buffer.

51

Experiment 2: Measure the Accuracy of Intruder Identification

Purpose: investigate the impacts of host mobility, number of attackers ,and connection scenarios on the detection accuracy of RLR

Input parameters: number of independent attackers, number of connections, host pause time

Output parameters: false positive alarm ratio, false negative alarm ratio

Observation: The increase in connections may improve the detection accuracy of RLR. When multiple attackers exist in the network, RLR has a high false positive ratio.

52

Accuracy of RLR: Single Attacker

30 hosts, 25 connections 30 hosts, 50 connections

Host Pause time (sec)

# of normal hosts identify the attacker

# of normal hosts marked as malicious

# of normal hosts identify the attacker

# of normal hosts marked as malicious

0 24 0.22 29 2.2

10 25 0 29 1.4

20 24 0 25 1.1

30 28 0 29 1.1

40 24 0 29 0.6

50 24 0.07 29 1.1

60 24 0.07 24 1.0

The accuracy of RLR when there is only one attacker in the system

53

Experiment 3: Measure the Communication Overhead

Purpose: investigate the impacts of host mobility and connection scenarios on the overhead of RLR

Input parameters: number of connections, host pause time

Output parameters: control packet overhead

Observation: When no false destination sequence attacks exist in the network, RLR introduces small packet overhead into the system.

54

X-axis is host pause time, which evaluates the mobility of host. Y-axis is normalized overhead (# of control packet / # of delivered data packet). 25 connections and 50 connections are considered. RLR increases the overhead slightly.

Control Packet Overhead

55

Research Opportunities: Improve Robustness of RLR

• Protect the good hosts from being framed by malicious hosts• The malicious hosts can frame the good hosts

by putting them into blacklist. • By lowering the trust values of both complainer

and complainee, we can restrict the impacts of the gossip distributed by the attackers.

• Adopting the trust management scheme proposed by Aberer and Despotovic [CIKM’01] to determine the lowering speed.

56

• Avoid putting every host into blacklist• Combining the host density and movement

model, we can estimate the time ratio that two hosts are neighbors

• The counter for a suspicious host decreases as time passes

• Adjusting the decreasing ratio to control the average percentage of time that a host stays in the blacklist of another host

57

• Defend against coordinated attacks• The behaviors of collusive attackers show

Byzantine manners. The malicious hosts may establish trust to frame other hosts, or conduct attacks alternatively to avoid being identified.

• Look for the effective methods to defend against such attacks. Possible research directions include:

• Apply classification methods to detect the hosts that have similar behavior patterns

• Study the behavior histories of the hosts that belong to the same group and detect the pattern of malicious behavior (time-based, order-based)

58

Conclusions on Intruder Identification• False destination sequence attacks can be

detected by the anomaly patterns of the sequence numbers

• Reverse labeling method can reconstruct the false routing tree

• Isolating the attackers brings a sharp increase in network performance

• On going research will improve the robustness of the mechanism and the accuracy of identification

59

Related Ongoing Research

A. Detecting wormhole attacks

B. Position-based private routing in ad hoc networks

C. Fault tolerant authentication in movable base station systems

D. Congestion avoidance routing in ad hoc networks