Download Our Inoculation Strategies for Victims of Viruses and the Sum

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Envy-free cake-cutting wikipedia , lookup

Nash equilibrium wikipedia , lookup

John Forbes Nash Jr. wikipedia , lookup

Chicken (game) wikipedia , lookup

Minimax wikipedia , lookup

Transcript
Our Inoculation Strategies for Victims of Viruses and the Sum-ofSquares Partition Problem*
Yafet
Farnaz
Zelalem
Spring 2017
*Aspnes, James, Kevin Chang, and Aleksandr Yampolskiy. "Inoculation strategies for victims of viruses and the sum-of-squares partition
Algorithm
Courses
University
problem." Journal
of Computer
and SystemEmory
Sciences
72.6 (2006): 1077-1093.
Outline
▶
▶
▶
▶
▶
▶
Motivation
Problem definition
Model
Nash Equilibria
Optimization
Sum of Squares problem
Algorithm Courses
Emory University
22
Motivation
Question: is it worth to install an antivirus software?
Kaspersky 2017 = $59.99
Value of data = $250, Probability of infection = 10%
Expected loss = $25
Answer: probably not
Algorithm Courses
Emory University
33
Problem definition (Cont.)
▶ Consider n (nodes in our graph) machines connected by a network, each
machine may or may not be protected
▶ Any virus that infects a machine also infects its unprotected neighbours
▶ Suppose that protecting a machine by installing an antivirus software costs the
owner of the machine C, but having the same machine be infected cost L (may
or may not be greater than C)
▶ Owner vs Society: If C > L no owner will install the antivirus, affecting the
whole society, selfish owner
Algorithm Courses
Emory University
44
Benevolent Dictator
▶
The Benevolent Dictator: decides which computers install an antivirus
▶
▶
selective inoculation of a few central machines can limit the spread of the virus
Questions to be addressed:
▶ How much of an improvement a centralized solution can provide?
▶ How easy is it to find a good centralized solution
Algorithm Courses
Emory University
5
Outline
▶
▶
▶
▶
▶
▶
Motivation
Problem definition
Model
Nash
Optimization
Sum Square problem
Algorithm Courses
Emory University
66
Model
▶ The network is an undirected graph G = (V,E).
▶ Installing anti-virus software is a one round noncooperative game.
▶ There are n player, the nodes of the graph, V =
{0,1,…,n-1}.
▶ Initially all nodes are not secure
Algorithm Courses
Emory University
77
Model: Strategies
▶ Each node has two options: do nothing or install the
antivirus
▶ ai = probability that node i installs anti-virus software
▶ The strategies of all n players can be given as a vector
Algorithm Courses
Emory University
88
Model: Strategies (Cont.)
▶ If ai is 0 or 1, node i adopts a pure strategy
▶ If 0<ai<1, node I adopts a mixed strategy
▶ Ia, set of nodes that have installed antivirus, secure
nodes
Algorithm Courses
Emory University
99
Model: Attack Model
▶
After all the nodes choose their strategies, an adversary
picks a starting point for infection uniformly at random
▶
A node i gets infected if it has no anti-virus software
installed and if any of its neighbors become infected
Algorithm Courses
Emory University
10
Model: Attack Model (Cont.)
▶
Example: Only node 3 installs anti-virus software.
Adversary chooses to infect node 2.
Algorithm Courses
Emory University
11
Model: Attack Graph
▶
Attack Graph Ga: subgraph of G induced by removing
secure nodes and their edges
network graph G
Algorithm Courses
Emory University
attack graph Ga= G - Ia
12
Model: Individual Costs
▶
Anti-virus software costs C. Loss from infection is L.
▶
For simplicity, assume C and L are known and the same
for all nodes
Cost of mixed strategy to node i:
▶
▶
Here, pi( ) = Pr[i is infected | i does not install an
antivirus]
Algorithm Courses
Emory University
Model: Social Cost
▶
Social cost of
Algorithm Courses
is simply a sum of individual costs:
Emory University
14
Outline
▶
▶
▶
▶
▶
▶
Motivation
Problem definition
Model
Nash Equilibria
Optimization
Sum Square problem
Algorithm Courses
Emory University
15 15
Strategies
▶ Two
▶
scenarios
each node decides individually
▶ Nash
▶
equilibrium
centralized plan
▶
dictator computes plan
Algorithm Courses
Emory University
Nash Equilibrium
▶
A stable state of a system involving the
interaction of different participants in
which:
no participant can gain by a unilateral change
of strategy
▶ if the strategies of the others remain
unchanged.
▶
Algorithm Courses
Emory University
Nash Equilibrium
▶ Assumptions
for our problem:
nodes are aware of each other’s choices
▶ nodes can easily evaluate C and L
▶ Nodes can compute Nash equilibrium in a
reasonable amount of time
▶
Algorithm Courses
Emory University
Nash Strategies: Characterization
There is a threshold t=Cn/L such that each node in a Nash
equilibrium
▶ will
install an anti-virus if it would otherwise end up in a
component of expected size > t in the attack graph.
▶ will not install an anti-virus if it would end up in a component
of expected size < t in the attack graph.
▶ is indifferent between installing and not installing when the
expected size = t in the attack graph.
Algorithm Courses
Emory University
Mixed Nash
S(i) expected size of insecure component
▶ A profile is in Nash equilibrium, iff:
▶
▶ For
all i, ai=1, S(i)>=t
▶ For all i, ai=0, S(i)<=t
▶ For all i, 0<ai<1, S(i)=t
Algorithm Courses
Emory University
Pure Nash
A simplified version with no random choices
▶ A profile is in Nash equilibrium, iff:
▶
▶ Every
component in attack graph has size at most t
▶ Inserting any secure node yields a component of size at least t
Algorithm Courses
Emory University
Nash Strategies
▶ Example:
▶
Then
Let C=0.5,L=1 so that t= Cn/L= 3.
is not a Nash equilibrium.
attack graph Ga= G - Ia
network graph G
0
1
0
1
2
3
2
3
4
Algorithm Courses
5
Emory University
4
5
Nash Strategies
▶ Suppose ai=0,
node i expects that it will lie in component of size greater
than t=Cn/L in attack graph
▶ Prob[i infected] = t/n.
▶
▶ Then
its expected loss from infection is >L·(Cn/L)/n = C
and it will switch to ai = 1.
Algorithm Courses
Emory University
▶
▶
It is NP-hard to compute a pure Nash equilibrium with
lowest and highest cost.
There exists a pure Nash equilibrium, which can be
achieved by a distributed, iterative process in 2n steps.
Algorithm Courses
Emory University
Nash strategies
Nash equilibrium can be computed in time O(n3)
▶ Using DFS we can test each node once.
▶A
▶ Compute
▶
the size of each component
O(|V| +|E|)=O(n2)
▶ 2n. n2=
Algorithm Courses
O(n3)
Emory University
Price of Anarchy
▶ Price
of anarchy measures how far away a Nash
equilibrium can be from the social optimum
▶ Formally, it is the worst-case ratio between cost of Nash
equilibrium and cost of social optimum
Algorithm Courses
Emory University
Price of Anarchy
▶ For
network G and costs C, L, we denote it:
Algorithm Courses
Emory University
A lower bound
Consider a star graph K1,n.
Let C=L(n-1)/n so that t=n-1.
1
n-1
2
n-2
3
0
…
G = K1,n
Algorithm Courses
Emory University
Price of Anarchy (cont.)
Then,
C+L(n-1)/n.
is an optimum strategy with cost
1
1
n-1
n-1
2
n-2
3
n-2
3
0
0
…
…
G = K1,n
Algorithm Courses
2
Emory University
Ga*
Price of Anarchy (cont.)
Then,
C+L(n-1)2/n.
is an optimum strategy with cost
1
n-1
1
2
n-2
n-1
3
n-2
3
0
0
…
…
G = K1,n
Algorithm Courses
2
Emory University
Ga*
Price of Anarchy
▶ Therefore,
▶
lower bound price of anarchy is:
(G, C, L)= L(n-1)/(2L(n-1)/n) = n/2
Algorithm Courses
Emory University
Price of Anarchy
▶ The
▶
▶
upper bound price of anarchy is:
(G, C, L) <= Ln/n=n.
(G, C, L) <=Cn/C=n
Algorithm Courses
Emory University
If C>L
if C<=L
Lower Bound: For a star graph K1,n,
(G, C, L) = n/2.
Upper Bound: For any graph G and any C, L,
(G, C, L)≤ n.
Price of anarchy in the game is
(G, C, L) = (n).
Algorithm Courses
Emory University
Outline
▶
▶
▶
▶
▶
▶
Motivation
Problem definition
Model
Nash Equilibria
Optimization
Sum Square problem
Algorithm Courses
Emory University
34 34
Optimization
▶
So, allowing users to selfishly choose whether or not to
install anti-virus software may be very inefficient
▶
Instead, let’s have a benevolent dictator compute and
impose a solution maximizing overall welfare
Algorithm Courses
Emory University
35 35
▶
Unfortunately,
Theorem: It is NP-hard to compute an optimum strategy
Fortunately,
Theorem: a strategy with cost at most O(log1.5 n)
OPT
can be computed in polytime
▶
Algorithm Courses
Emory University
36
Outline
▶
▶
▶
▶
▶
▶
Motivation
Problem definition
Model
Nash
Optimization
Sum Square problem
Algorithm Courses
Emory University
37 37
Sum of Squares Problem
Total Cost = C*(number of secure nodes)+L* Pr(infected node v) =
C |m| + (L/n)Σ Ki 2
Assumption: The number of node (m) is known
The network security problem is reduced to find the best
solution to the problem of removing m nodes to minimize
the sum of square.
Algorithm Courses
Emory University
38 38
Sum of square partition problem
NP-hard
Given a graph G = (V, E), by removing a set of at most m
nodes, partition the graph into mutually disconnected
components {Hi}, such that ∑|Hi|2 is minimum.
(𝝰,𝝱)-bicriterion approximation algorithm:
The output is a node cut consisting of at most 𝝰m nodes and
partitioning graph into connectd components {Hi}, such that ∑|Hi|2 <=
𝝱 OPT
Algorithm Courses
Emory University
39
Theorem
There exists a
polynomial time (O(log1.5n), O(1))-bicriterion approximation algorithm
for the sum-of- squares partition problem
OPT: The cost of the optimum solution for the inoculation
problem,then there exists a polynomial time approximation
algorithm that finds a solution with cost at most
O(log1.5n)OPT
Algorithm Courses
Emory University
40 40
Sparsest Cut
Sparsest Cut
S
S’
C-balanced separator
Both NP-hard
Algorithm Courses
Emory University
41
Arora-Rao-Vazirani Algorithm*
S. Arora, S. Rao, and U. Vazirani. Expander flows, geometric embeddings and graph partitioning.36th Annual ACM Symposium on the Theory of Computing,
Exists O(log 0.5 n)-approximation for the following
problem
Given graph G=(V,E), find a node cut that partitions the nodes
of G into three sets:
▶ Two disconnected subgraphs with nodes sets V1 and V2
▶ A set of removed nodes R, such that the following quantity is
maximized.
Sparsity of the Cut
Algorithm Courses
Emory University
42 42
Cost-Effectiveness
Connected subgraph H with k nodes that is split into sets
V1 and V2, and R then the cut’s cost-effectiveness is:
Algorithm Courses
Emory University
43 43
Cost-Effectiveness vs. Sparsity
▶
Sparsity
▶
Cost effectiveness
𝝱
𝝰
𝝰 > 2𝝱
Cost effectivenss > 2 sparsity
Algorithm Courses
Emory University
44 44
Algorithm
Input: G = Graph(V,E)
M>0
Find the approximate most cos-effective cut in
each connected component
G
Partition the graph G into H1,H2 by removing
a sparse node cut.
If the sparse cut removes at most (20c log 𝑛 )
m nodes in H1 … Hk, continues, if there is not
such component halts the process.
Algorithm Courses
Emory University
H1
45
H2
Algorithm
Choose the component Hj among those components considered in previous
step for which the cost-effectiveness is highest.
Set F = F U R , if |F| > (36 log 1.5 n)m nodes then halt.
Otherwise, repeat.
H1
Algorithm Courses
H2
Emory University
Runtime
▶ Algorithm
halts as soon as we augment the set of marked
nodes |F| >(36c log1.5 n)m
▶ At the beginning of each iteration, F contains at most (36c
log1.5 n)m marked nodes.
▶ Since we add at most (20c log0.5 n)m marked nodes per
iteration.
Algorithm Courses
Emory University
Algorithm Courses
Emory University
48