On the Evolution of Adversary Models
(from the beginning to sensor networks)
Virgil D. Gligor
Electrical and Computer Engineering
University of Maryland
College Park, MD. 20742
gligor@umd.edu
Lisbon, Portugal
July 17-18, 2007
VDG, July 17, 2007
Copyright © 2007
1
Overview
1. New Technologies often require a New Adversary Def.
- continuous state of vulnerability
2. Why is the New Adversary Different ?
- ex. sensor, mesh networks, MANETs
- countermeasures
3. Challenge: find “good enough” security countermeasures
4. Proposal: Information Assurance Institute
VDG, July 17, 2007
Copyright © 2007
2
A system without an adversary definition cannot
possibly be insecure; it can only be astonishing…
… astonishment is a much underrated security vice.
(Principle of Least Astonishment)
VDG, July 17, 2007
Copyright © 2007
3
Why an Adv. Def. is a fundamental concern ?
1. New Technology > Vulnerability ~> Adversary <~> Methods & Tools
-sharing user-mode
programs& data;
- computing utility
(early – mid 1960s)
confidentiality and
integrity breaches;
system penetration;
- shared stateful
DoS instances
services
e.g., DBMS, net. protocols
dyn. resource alloc.
(early - mid 1970s)
untrusted usermode programs
& subsystems
sys. vs. user mode (’62->)
rings, sec. kernel (’65, ‘72)
FHM (’75) theory/tool (’91)*
acc. policy models (’71)
untrusted user
processes;
concurrent,
coord. attacks
DoS = a diff. prob.(83-’85)*
formal spec. & verif. (’88)*
DoS models (’92 -> )
- PCs, LANs;
read, modify, block,
public-domain Crypto replay, forge
(mid 1970s)
messages
“man in the middle” informal: NS, DS (’78–81)
active, adaptive
semi-formal: DY (‘83)
network adversary Byzantine (‘82 –>)
crypto attk models (‘84->)
auth. prot. analysis (87->)
- internetworking
(mid – late 1980s)
geo. distributed,
coordinated
attacks
large-scale effects:
worms, viruses,
DDoS (e.g., flooding)
virus scans, tracebacks
intrusion detection
(mid ’90s ->)
2. Technology Cost -> 0, Security Concerns persist
VDG, July 17, 2007
Copyright © 2007
4
Continuous State of Vulnerability
New
New
Technology > Vulnerability ~>
+/- O(months)
New
New Analysis
Adversary Model <~> Method & Tools
+O(years)
+O(years)
… a perennial challenge (“fighting old wars”)
New
Technology ~>
New
Vulnerability
Old
Adversary Model
Reuse of Old
(Secure)
Systems &
Protocols
mismatch
VDG, July 17, 2007
Copyright © 2007
5
New Technology Ex.: Sensor Networks
Claim
Sensor Networks introduce:
- new, unique vulnerabilities: nodes captured and replicated
- new adversary: different from and Dolev-Yao and traditional Byzantine adv.s
and
- require new methods and tools: emergent algorithms & properties
(for imperfect but good-enough security)
Mesh Networks have similar but not identical characteristics
VDG, July 17, 2007
Copyright © 2007
6
Limited Physical Node Protection
Two Extreme Examples
Low end: Smart Cards (< $15)
High end: IBM 4764 co-proc. (~ $9K)
- no tamper resistance
- non-invasive phys. attacks
- side-channel (timing, DPA)
- tamper resistance, real time response
- independent battery, secure clock
- battery-backed RAM (BBRAM)
- wrapping: several layers of non-metallic
- unusual operating conditions
- temperature, power clock glitches grid of conductors in a grounded shield
to reduce detectable EM emanations
- invasive phys. attacks
- tamper detection sensors (+ battery)
- chip removal from plastic cover
- temp., humidity, pressure, voltage,
- microprobes, electron beams
clock, ionizing radiation
- response: erase BBRAM, reset device
VDG, July 17, 2007
Copyright © 2007
7
Limited Physical Node Protection
Observation:
a single on-chip secret key is sufficient to protect
(e.g., via Authenticated Encryption)
many other memory-stored secrets (e.g., node keys)
Problem:
how do we protect that single on-chip secret key ?
Potential Solution: Physically Unclonable Functions (PUFs)
observation: each IC has unique timing
basic PUF: Challenge extracts unique, secret
Response (i.e., secret key) from
IC-hidden, unique timing sequence
VDG, July 17, 2007
Copyright © 2007
8
Basic PUF circuit [Jae W. Lee et al. VLSI ‘04]
unknown challenge bit
IC
b62
feed-fwd arbiter
255
0
1
Arbiter
b0
b1
bi=0
switch
VDG, July 17, 2007
b2
b61
b128
LFSR
Response
e.g., 255 bits
Challenge
e.g., 128 bits
Arbiter
0
Arbiter
1
bi=1
Arbiter operation
Copyright © 2007
9
Basic PUF circuit [Jae W. Lee et al. VLSI ‘04]
Basic PUF counters:
brute-force attacks (2*128 challenge-response pairs => impractical)
duplication (different timing => different Secret Response)
invasive attacks (timing modification => different Secret Response)
However,
Pr. 1: adversary can build timing model of Arbiter’s output
=> can build clone for secret-key generation
Pr. 2: Arbiter’ output (i.e., secret-key generation) is unreliable
Reality: intra-chip timing variation (e.g., temp, pressure, voltage)
=> errors in Arbiter’s output (e.g., max. error: 4 – 9%)
VDG, July 17, 2007
Copyright © 2007
10
Suggested PUF circuit [Ed Suh et al. ISCA ‘05]
Solution to Pr. 1:
hash Arbiter’s output to provide new Response
- cannot discover Arbiter output from known Challenges and new Responses
Solution to Pr. 2:
add Error Correcting Codes (ECCs) on Arbiter’s output
e.g., use BCH(n, k, d)
n(timing bits) = k(secret bits) + b(syndrome bits) for (d-1)/2 errors
BCH (255,63,61)
=> 30 (> 10%n > max. no.) errors in Arbiter’s
output are corrected
> 30 errors ? (probability is 2.4 X10-6)
probability of incorrect output is smaller but not zero
hash Arbiter’s output and verify against stored Hash(Response)
VDG, July 17, 2007
Copyright © 2007
11
Suggested PUF circuit
IC
BCH
b62
feed-fwd arbiter
255
bits
Arbiter
b0
b1
b2
b61
b128
Hash
LFSR
known
Syndrome
e.g., 192 bits
secret
Response
known
Challenge
e.g., 128 bit
generate response: C -> R, S; retrieve response: C, S -> R
However, syndrome reveals some (e.g., b=192) bits of Arbiter’s output (n=255)
(Off-line) Verifiable-Plaintext Attack:
Get C, S, hash(R); guess remaining (e.g., 63) bits of Arbiter’s output; verify new R;
repeat verifiable guesses until Arbiter’s output is known; discover secret key
VDG, July 17, 2007
Copyright © 2007
12
Some Characteristics of Sensor Networks
1. Ease of Network Deployment and Extension
- scalability => simply drop sensors at desired locations
- key connectivity via key pre-distribution =>
neither administrative intervention nor TTP interaction
2. Low Cost, Commodity Hardware
- low cost => physical node shielding is impractical
=> ease of access to internal node state
(Q: how good should physical node shielding be to prevent access
to a sensor’s internal state ?)
3. Unattended Node Operation in Hostile Areas =>
adversary can capture, replicate nodes (and node states)
VDG, July 17, 2007
Copyright © 2007
13
Replicated Node Insertion: How Easy ?
NEIGHBORHOOD j
NEIGHBORHOOD i
1
NEIGHBORHOOD k
i
3
2
VDG, July 17, 2007
Copyright © 2007
14
Attack Coordination among Replicas:
How Easy ?
NEIGHBORHOOD j
NEIGHBORHOOD i
3
1
NEIGHBORHOOD k
3
collusion
i
3
2
Note: Replica IDs are cryptographically bound to pre-distributed keys and cannot be changed
VDG, July 17, 2007
Copyright © 2007
15
New vs. Old Adversary
Old (Dolev-Yao) Adversary can
- control network operation
- man-in-the-middle: read, replay, forge, block, modify, insert messages
anywhere in the network
- send/receive any message to/from any legitimate principal (e.g., node)
- act as a legitimate principal of the network
Old (Dolev-Yao) Adversary cannot
1) adaptively capture legitimate principals’ nodes and discover a legitimate
principal’s secrets
2) adaptively modify network and trust topology (e.g., by node replication)
Old Byzantine Adversaries
- can do 1) but not 2)
- consensus problems impose fixed thresholds for captured nodes
(e.g., t < n/2, t < n/3) and fixed number of nodes, n.
VDG, July 17, 2007
Copyright © 2007
16
Countermeasures for Handling New Adv.?
1. Detection and Recovery
-
Ex. Detection of node-replica attacks
Cost ? Traditional vs. Emergent Protocols
Advantage: always possible, good enough detection
Disadvantage: damage possible before detection
2. Avoidance: early detection of adversary’s presence
-
Ex. Periodic monitoring
- Cost vs. timely detection ? False negatives/positives ?
- Advantage: avoids damage done by new adversary
- Disadvantage: not always practical in MANETs, sensor and mesh
networks
3. Prevention: survive attacks by “privileged insiders”
VDG, July 17, 2007
Ex. Subsystems that survive administrators’ attacks (e.g., auth)
Cost vs. design credibility ? Manifest correctness
Advantage: prevent damage; Disadvantage: very limited use
Copyright © 2007
17
Example of Detection and Recovery
(IEEE S&P, May 2005)
- naïve: each node broadcasts <ID, “locator,” signature>
perfect replica detection: ID collisions, different Iocators
complexity: O(n2) messages
- realistic: each node broadcasts locally <ID, “locator,” signature>
local neighbors further broadcast to g << n random witnesses
good enough replica detection: ID collision, different
Iocators at witness
detection probability: 70 - 80% is good enough
complexity: O(n x sqrt(n)) messages
VDG, July 17, 2007
Copyright © 2007
18
A New App.: Distributed Sensing
VDG, July 17, 2007
Copyright © 2007
19
A New Application: Distributed Sensing
Application: a set of m sensors observe and signal an event
- each sensor broadcasts “1” whenever it senses the event;
else, it does nothing
- if t  m broadcasts, all m sensors signal event to neighbors; else do nothing
Operational Constraints
- absence of event cannot be sensed (e.g., no periodic “0” broadcasts)
- broadcasts are reliable and synchronous (i.e., counted in sessions)
Adversary Goals: violate integrity (i.e., issues t  m/2 false broadcasts)
deny service (i.e., t > m/2, suppresses m-t+1 broadcasts)
New (Distributed-Sensing) Adversary
- captures nodes, forges, replays or suppresses (jams) broadcasts
(within same or across different sessions)
- increases broadcast count with outsiders’ false broadcasts
VDG, July 17, 2007
Copyright © 2007
20
An Example: distributed revocation decision
[IEEE TDSC, Sept. 2005]
m=6, t = 4 votes in a session => revoke target
Keying
Neighborhood
revocation
target
Communication
Neighborhood
4
10
3
2
8
14
5
11
1
7
12
VDG, July 17, 2007
13
9
6
Copyright © 2007
21
New vs. Old Adversary
A (Reactive) Byzantine Agreement Problem ?
- both global event and its absence are (“1/0”) broadcast by each node
- strong constraint on t ; i.e., no PKI => t > 2/3m; PKI => t >m/2
- fixed, known group membership
No.
New (Distributed-Sensing) Adv. =/= Old (Byzantine) Adv.
- new adversary need not forge, initiate, or replay “0” broadcasts
- new adversary’s strength depends on a weaker t (e.g., t < m/2)
- new adversary may modify membership to increase broadcast count ( > t)
VDG, July 17, 2007
Copyright © 2007
22
Conclusions
1. New Technologies => New Adversary Definitions
- avoid “fighting the last war”
- security is a fundamental concern of IT
2. No single method of countering new and powerful adversaries
- detection
- avoidance (current focus)
- prevention (future)
3. How effective are the countermeasures ?
- provide “good enough” security; e.g., probabilistic security
properties
VDG, July 17, 2007
Copyright © 2007
23