Transcript Investigating the impact of real-world factors on Internet worm

Investigating the Impact of Real-World Factors
on Internet Worm Propagation
Daniel Ray, Charles Ward, Bogdan Munteanu, Jonathan Blackwell,
Xiaoyan Hong,
Computer Science
University of Alabama
Jun Li,
Department of Computer and Information Science
University of Oregon
Worms: Why Do We Care?
 Internet worms have been costly and destructive
•
•
•
•
Traffic causes network degradation
Infected hosts are often unusable
Repair is labor-intensive
Code Red v.2 and Slammer are estimated to cost $2bn
and $1bn, respectively
 Fast self-propagating worms
• Slammer reached its peak infection rate in ~3 minutes
• Infected ≥90% vulnerable hosts in ≤10 min
• Theoretical maximal speed: ~15 seconds
• 15 seconds to infect most vulnerable hosts
2
Analyzing Worms
 Analytical methods
• Based on mathematical epidemiological models
• Easy to compute, with limitations
• Model may or may not represent that of the real-world
• Some margin of error is to be expected
• Also very rigid/Inflexible
 Empirical model
• build components that act like real world components.
• Coarse Granularity
– abstract out individual packets or even individual nodes
• Fine Granularity
– Have components that simulate all elements of the network down to and
including individual packets
• Is where our research fits in
3
What Do We Study?
 The impact of real-world factors on Internet worm
propagation
 Factors we focus on:
• IP address allocation strategy
• Worm scanning methods
• Wireless media
 Use packet-Level network simulator: GTNetS
4
Wireless Internet
 Wireless networks
•
•
•
•
WLAN
Mobile ad hoc network
Multihop mesh wireless network
Vehicular networks
 Possible influences
•
•
•
•
WLAN address allocation
Bandwidth
Use behavior – connectivity
Device vulnerability – e.g. Bluetooth
5
Previous Work (a few examples)
 Chen, Gao, Kwiat, “Modeling The Spread of Active Worms”
• give analytical model entitled Analytical Active Worm Propagation (AAWP)
• do not deal specifically with connection type or network topology.
• using a variant of the Code Red & NIMDA Worm, which are TCP worms
 Wei, Mirkovic, Swany, “Distributed Worm Simulation with a Realistic Internet Model”
•
•
•
•
examine worms using a similar but less flexible packet level simulator
UDP worms (TCP be approximated), random and subnet scanning
network topology at AS level
adjusting ratio of live hosts in address space for each AS
 Weaver, Staniford, Paxson, “Very Fast Containment of Scanning Worms”
• employ a simulator to test worm retardation algorithm by starting with an
algorithm for containment of scanning worms
• focus on scanning in general, rather than specific scanning types or
connection types
• use a probability method to determine if an attack is likely to succeed based
on the expected amount of non-attack traffic
• LAN or company workstation networks instead of internet like topology.
• address space adjustable through the likelihood of successful infection.
6
Previous Work (cont’d)
 GTNetS folks: Riley, Sharif, and Lee
• “Large-Scale Network Simulations with GTNetS”, “Simulating Internet
Worms”
• GTNetS design to model networks
• GTNetS capabilities of modeling worms
• Investigated:
• randomly scanning TCP worms
– TCP payload size, # of Parallel TCP Connections
• randomly scanning UDP worms
– Length bandwidth, Scan rate, payload size
 More work
• Self-learning worm using importance scan
• Self-stopping worms
• Defending hit-list worm using address space randomization
7
Previous Work: Wireless (cont’d)
 Khayam, Radha, VANET 04
•
•
•
•
warm spread over ad hoc vehicular network
SIR (susceptible, Infected and removed) epidemic model
Network: a new geometric random graph
Impact: vehicle traffic density – average node degree used
 Hoh, Gruteser, WSPWN06
• Infection may be limited due to device diversity.
• Propagate rate and infection rate.
• Experiment:
• Southern New Jersey highway network
• SIR model, traffic simulator PARAMICS
– 10min reach 11.6 km, 75m/s if 5% of vehicles are susceptible
– Slower but still fast enough to make containment difficulty
 Worms in wireless sensors (analytical models)
 Not sufficient work on detailed empirical analysis
8
Why GTNetS
 The simulator we chose to use to facilitate our research
• Fully Functional, Fully Adaptable, Packet-Level Network Simulator
 Has a worm packet class which is fully extensible
 Allows the simulation to handle worm characteristics
• Support TCP or UDP connections
• Varying infection lengths, infection ports, scan rate (UDP) and
number of connections (TCP)
• Allows for varying IP block scanning methods
 Network topology support (but weak for our purpose)
• Simple network structure: Star, Dumbbell, Trees
• Interfaces support BRITE network simulator
• to generate internet like topologies.
9
Factors Currently Studying
 Topology (IP address allocation)
• Dense vs Sparse
• IPv6: The ratio of active simulated nodes in the address space can be
limited in such a fashion to mimic the distribution of nodes in the early
stages of IPv6.
• Internet-like topology vs other topology (deeper tree or wider tree)
• Wireless LAN address allocation
10
Worm Scanning Methods
 IP address block scanning:
• Random Scan
• Local Preference Scan
• Hit-List Scan
 Connection types, worms at the packet level
• UDP: they are faster, more effective
• TCP: TCP connections can increase the effectiveness of worm scan
over the cost of TCP overhead.
 Note: hit-list was the most likely to be effected positively.
11
Preliminary GTNetS Simulation
 Network topology:
• Internet like,
• Addresses are chosen randomly and assigned to topology randomly
• IP address space population density
– Sparse (IPv4 like): 1/35 Addresses in the space are occupied
– Dense (IPv6 like): 1/135 Addresses in the space are occupied
• Synthetic topology
• Wide tree: backbone + local WLANs
• Deep tree: more administration penetration
 Worm IP block Scanning method
• UDP worms
• uniform random and local preference based on examples
• hit-list worm with local preference scanning
• TCP worm: hit-list worm
• Port scanning are not used
12
Preliminary GTNetS Simulation (cont’d)
 Network constants
• Size of network
• No other network traffic
• Can effect worm spread but,
• Largely a function of the topology
• Difficult to simulate real-world situation
• Individual node vulnerability
 Worm constants
• Scan rate/number of TCP connections
• Infection Length
 Each simulation was run until all vulnerable nodes were infected or until
computer memory was consumed.
13
Worm Types: Uniform Random vs. Local Pref
Uniform random
dense
sparse
Local preference
dense
sparse
 universally quicker on
dense networks
14
Worm Types: TCP Hit-List vs. UDP Hit-List
TCP hit-list
dense
sparse
UDP hit-list
dense
sparse
 TCP causes a lot of overhead
but no gain in speed
 Local Preference and Hit-List
• Worse than uniform random
on dense graphs
• Better than uniform random on
sparse graphs
15
Dense and Sparse Graphs
Dense net



Blue – Uniform Random
Red – Hit-list
Green - Local Pref
Sparse net

Worms spread trend similar
•
•

Local-pref slower than hit-list
Uniform random shifted
regardless of worm type, sparse
networks retard spread
16
Low Bandwidth Wireless Links
Uniform
Wider-tree
deeper-tree
Local preference
Wider-tree
deeper-tree
 TCP worm
 Uniform vs. local preference.
 100% allocated space
17
Low Bandwidth Wireless Links (cont’d)
Uniform random
Wider-tree
deeper-tree
Local preference
Wider-tree
 UDP worm
 Uniform vs. local preference.
deeper-tree
 100% allocated space
18
Summary…
 Impact of real-world factors on Internet worm propagation
 Factors discussed:
• IP address allocation strategy: dense, sparse, wider tree, deeper
• Worm scanning methods: uniform random, hit-list, local preference
• Wireless media: low-bandwidth in two topology
 Future work:
 More worm scanning types, e.g.,
• Permutation scanning, topological scanning
• Hit-List with other scanning methods
• Emerging ones
 Influence from other network traffic
 More topology testing, including wireless network.
19
Questions?
Thanks!