A Numerical Study on the Wiretap Network with a Simple
Network Topology
Fan Cheng and Vincent Tan
Department of Electrical and Computer Engineering National University of Singapore
Mathematical Tools of Information-Theoretic Security Workshop September, 23 – 25, 2015
Problem Statement
Source Sink
Wiretappers 1:
2:
Problem Statement
Source Sink
Wiretappers 1:
2:
Problem Statement
Wiretap Network Strategy
• Source
• M: message
• K: random key, I(M;K) =0
• Intermediate: mix the ciphertext
• Sink: recover M
• A: no information about M
Problem Statement
Wiretap Network Strategy
• Source
• M: message
• K: random key, I(M;K) =0
• Intermediate: mix the ciphertext
• Sink: recover M
• A: no information about M
Goal
• Maximize H
(
M)
• Minimize H
(
K)
Problem Statement
Wiretap Network Strategy
• Source
• M: message
• K: random key, I(M;K) =0
• Intermediate: mix the ciphertext
• Sink: recover M
• A: no information about M
Goal
• Maximize H
(
M)
• Minimize H
(
K)
Very hard!
Problem Statement
Our Model
S T R
X1
X2
X3
X4
X5
X6
Problem Statement
Our Model
S T R
X1
X2
X3
X4
X5
X6
Objective:
For a given wiretap pattern
A
:τ
A=
minHH((MK))Problem Statement
Our Model
S T R
X1
X2
X3
X4
X5
X6
Objective:
For a given wiretap pattern
A
:τ
A=
minHH((MK))Question:
Is routing optimal?
Outline
• Literature review
• Problem formulation and methods
• Results and open problems
Wiretap Network
• N. Cai and R. W. Yeung, “Secure network coding,” 2002 IEEE International Symposium on Information Theory.
• N. Cai and R. W. Yeung, “Secure Network Coding on a Wiretap Network,” IEEE Transactions on Information Theory, vol. 57, no.
1, pp. 424 - 435, Jan. 2011.
Sinks Sources
Wiretap network (
G = (V, E), S, U , A
)Wiretapper, Wiretap Set, Wiretap Pattern
S T R
X1
X2
X3
X4
X5
X6
Wiretapper
→
Wiretap Set→
Wiretapper PatternA
, wiretap pattern:A = {
A1,
A2, ..., }
, where Ai⊆ E
Example
Three Wiretappers with wiretap sets A1
= {
X1,
X2}
, A2= {
X2,
X3,
X5}
, A3= {
X1,
X2,
X4,
X6}
.A = {
A1,
A2,
A3}
.(
H(
M),
H(
K))
? whenA
is given.63 subsets of
E →
263wiretap patterns.Wiretapper, Wiretap Set, Wiretap Pattern
S T R
X1
X2
X3
X4
X5
X6
Wiretapper
→
Wiretap Set→
Wiretapper PatternA
, wiretap pattern:A = {
A1,
A2, ..., }
, where Ai⊆ E Example
Three Wiretappers with wiretap sets A1
= {
X1,
X2}
, A2= {
X2,
X3,
X5}
, A3= {
X1,
X2,
X4,
X6}
.A = {
A1,
A2,
A3}
.(
H(
M),
H(
K))
? whenA
is given.63 subsets of
E →
263wiretap patterns.Wiretapper, Wiretap Set, Wiretap Pattern
S T R
X1
X2
X3
X4
X5
X6
Wiretapper
→
Wiretap Set→
Wiretapper PatternA
, wiretap pattern:A = {
A1,
A2, ..., }
, where Ai⊆ E Example
Three Wiretappers with wiretap sets A1
= {
X1,
X2}
, A2= {
X2,
X3,
X5}
, A3= {
X1,
X2,
X4,
X6}
.A = {
A1,
A2,
A3}
.(
H(
M),
H(
K))
? whenA
is given.63 subsets of
E →
263wiretap patterns.Existing Results on Wiretap Networks
• 1. Single source multiple sinks,
A = {
W: |
W| =
r,
W⊆ E}
:• H(M) ≤ (n−r)log q;
• H(K) ≥r/(n−r)H(M). (Cai and Yeung, 2010)
• 2. General multi-source multi-sink wiretap network
⇐⇒
General entropic region problem. (Chan and Grant, 2014)Open problem
Single source and single sink network witharbitrary
A
.Existing Results on Wiretap Networks
• 1. Single source multiple sinks,
A = {
W: |
W| =
r,
W⊆ E}
:• H(M) ≤ (n−r)log q;
• H(K) ≥r/(n−r)H(M). (Cai and Yeung, 2010)
• 2. General multi-source multi-sink wiretap network
⇐⇒
General entropic region problem. (Chan and Grant, 2014)Open problem
Single source and single sink network witharbitrary
A
.CutSet Bound
F. Cheng and R. W. Yeung,“Performance Bounds on a Wiretap Network withArbitraryWiretap Sets,” IEEE Trans. Inform. Theory, vol. 60, no.
6, pp. 3345-3358, Jun. 2014.
Graph Cut
S
T
x1
x2
xn−1 xn
Cutset bound
H
(
K)
H
(
M) ≥
maxn
X
i=1
xi
−
1 s.t. 0≤
xi≤
1,
1≤
i≤
nX
ej∈A
xj
≤
1, ∀
A∈ A
Assume H
(
K)
= 1Problem Formulation: a Simple Network
Network model
S T R
X1
X2 X3
X4
X5 X6
(3, 3) wiretap network
•
G = (V, E)
•
S = {
S}
,U = {
R}
•
A ⊆
2EEntropy Equations Network model
S T R
X1
X2
X3
X4
X5
X6
(3, 3) wiretap network
Source: I(
M;
K) =
0H
(
X1,
X2,
X3|
M,
K) =
0 Node T: H(
X4,
X5,
X6|
X1,
X2,
X3) =
0Sink: H
(
M|
X4,
X5,
X6) =
0 (Level-I) or H(
M,
K|
X4,
X5,
X6) =
0 (Level-II) Security: I(
XA;
M) =
0, ∀
A∈ A
Question
τ
A=
minH(
K)
H(
M)
For a fixedA
, is routing optimal?Numerical Study!
Lower and Upper Bounds on τ
A• Lower bounds: Cutset bound, Shannon Bound
• Upper bounds: Routing bound, Linear network coding bound Cutset
≤
Shannon≤ τ
A≤
Linear network coding≤
RoutingShannon Bound
• All information measures H
(·|·)
, I(·; ·|·)
are non-negative;H
(
X1|
X2),
H(
X1,
X2,
X3),
I(
X;
Y) ≥
0• Linearity exists among information measures;
H
(
X1,
X2) =
H(
X1|
X2) +
H(
X2)
• All information inequalities are linear; H
(
K) ≥
H(
M)
⇒
Idea: choose some asdecision variablesto represent the others.Shannon Cipher System
Message: X1, Key: X2, Ciphertext: X3:
X1
S R
X1 X2
X3
I
(
X1;
X2) =
0;
H(
X3|
X1,
X2) =
0;
H(
X1|
X2,
X3) =
0;
I(
X1;
X3) =
0 Objective: H(
X2) ≥
H(
X1)
(Perfect Secrecy Theorem)Shannon Bound
• All information measures H
(·|·)
, I(·; ·|·)
are non-negative;H
(
X1|
X2),
H(
X1,
X2,
X3),
I(
X;
Y) ≥
0• Linearity exists among information measures;
H
(
X1,
X2) =
H(
X1|
X2) +
H(
X2)
• All information inequalities are linear; H
(
K) ≥
H(
M)
⇒
Idea: choose some asdecision variablesto represent the others.Shannon Cipher System
Message: X1, Key: X2, Ciphertext: X3:
X1
S R
X1 X2
X3
I
(
X1;
X2) =
0;
H(
X3|
X1,
X2) =
0;
H(
X1|
X2,
X3) =
0;
I(
X1;
X3) =
0 Objective: H(
X2) ≥
H(
X1)
(Perfect Secrecy Theorem)Shannon Bound (cont’d)
• Decision variables: H
(
X1|
X2,
X3)
, H(
X2|
X1,
X3),
H(
X3|
X1,
X2)
, I(
X1;
X2),
I(
X1;
X2|
X3)
,I(
X1;
X3),
I(
X1;
X3|
X2)
,I(
X2;
X3),
I(
X2;
X3|
X1)
Denotez
:=
H
(
X1|
X2,
X3)
H(
X2|
X1,
X3)
. . .
I(
X2;
X3|
X1)
• H
(
X2) ≥
H(
X1)
⇔
cTz≥
0• I
(
X1;
X2) =
0;
H(
X3|
X1,
X2) =
0;
H(
X1|
X2,
X3) =
0;
I(
X1;
X3) =
0⇔
a11
. . . . . . . . . . . . . . . . . . . . . . . .
z≥
0⇔
Az≥
0•
min cTz
,
s.t. Az≥
0Shannon Bound (cont’d)
Minimal representation
Let
[
n] = {
1,
2, . . . ,
n}
. X1, X2,. . .
, Xn. The minimal decision variables:(i) H
(
Xi|
X[n]−{i}),
i∈ [
n]
;(ii) I
(
Xi;
Xj|
XK)
, where i6=
j and K⊆ [
n] − {
i,
j}
.⇒
Shannon bound:min cTz s.t. Az
≥
0Remark
• Always exists
• Hard to compute, hard to understand, optimality is unknown
• Automatical tools
Shannon Bound (cont’d)
Minimal representation
Let
[
n] = {
1,
2, . . . ,
n}
. X1, X2,. . .
, Xn. The minimal decision variables:(i) H
(
Xi|
X[n]−{i}),
i∈ [
n]
;(ii) I
(
Xi;
Xj|
XK)
, where i6=
j and K⊆ [
n] − {
i,
j}
.⇒
Shannon bound:min cTz s.t. Az
≥
0Remark
• Always exists
• Hard to compute, hard to understand, optimality is unknown
• Automatical tools
Shannon Bound: ITIP/Xitip
Information Theoretic Inequalities Prover
For n random variables, a linear program in n
+
n22n−2variables
Routing Bound and Linear Network Coding Bound
S T R
X1
X2 X3
X4
X5
X6
Routing bound
⇒
Special Cutset bound Linear network coding bound (LNCB): V1,
V2, ...,
Vm⊆ F
nqrank
(
V1)
,rank(
V1,
V2) →
entropy functions.rank
(
V1,
V2) ≤
rank(
V1) +
rank(
V2)
LNCB→
All the inequalities over V1,
V2, ...,
Vm• m
=
4, Ingleton• m
=
5, Dougherty, Freiling, Zeger, 28• m
=
6, unknown, at least 1 millionReduction of Wiretap Patterns Network model
S T R
X1
X2 X3
X4
X5 X6
A ⊆
2E, 226−1A1
= {
1,
2,
4},
A2= {
2,
4} ⇒
delete A2fromA A
: Ai6⊂
Aj(
i6=
j)
(antichain)|
E
| 0 1 2 3 4 5 6 7# 2 3 6 20 168 7581 7
.
8∗
106 2.
4∗
1012Numerical Study: Algorithm
Cut-set
≤
Shannon≤ τ
A≤
Linear Network Coding≤
RoutingAlgorithm for assessing the tightness of the routing bound
For each wiretap pattern,1. Compute the cut-set bounds on
(
S,
T)
and(
T,
R)
: l1. 2. Compute the routing bound on S→
T→
R: l2. 3. If l1==l2, thenτ
A=
l1. Proceed to Step 5.4. Compare l2with the Shannon bound in ITIP/Xitip: If equal, then
τ
A=
l2. Otherwise, agapis detected.5. Proceed to the next wiretap pattern.
Numerical Study: Result
• Gaps exist only if 4
≤ |A| ≤
12• For almost 80
%
of the wiretap patterns, cut-set bounds=
routing bounds• In the Level-I
(
3,
3)
network, there are around 159,
258 wiretap patterns (2%
of all the wiretap patterns) routing bounds6=
Shannon bounds
• In the Level-II
(
3,
3)
network, there are around 32,
472 wiretap patterns (0.4%
of all the wiretap patterns) routing bounds6=
Shannon bounds
Problem
How to fill this gap?
Cut-set
≤
Shannon≤ τ
A≤
Linear Network Coding≤
RoutingNumerical Study: Result
• Gaps exist only if 4
≤ |A| ≤
12• For almost 80
%
of the wiretap patterns, cut-set bounds=
routing bounds• In the Level-I
(
3,
3)
network, there are around 159,
258 wiretap patterns (2%
of all the wiretap patterns) routing bounds6=
Shannon bounds
• In the Level-II
(
3,
3)
network, there are around 32,
472 wiretap patterns (0.4%
of all the wiretap patterns) routing bounds6=
Shannon bounds
Problem
How to fill this gap?
Cut-set
≤
Shannon≤ τ
A≤
Linear Network Coding≤
RoutingCase Study: Routing Vs. Coding
S T R
X1
X2
X3
X4
X5
X6
A1
= {
2,
3,
5}
, A2= {
1,
4,
5}
, A3= {
1,
3,
6}
, and A4= {
2,
4,
6}
. Routing bound: 3. Both of the Shannon bounds (Level I/II): 2.X1
=
K1 X4=
M+
2K1+
2K2X2
=
K2 X5=
M+
K1+
2K2X3
=
M+
K1+
K2 X6=
M+
2K1+
K2Case Study: Routing Vs. Coding
S T R
X1
X2
X3
X4
X5
X6
A1
= {
2,
3,
5}
, A2= {
1,
4,
5}
, A3= {
1,
3,
6}
, and A4= {
2,
4,
6}
. Routing bound: 3. Both of the Shannon bounds (Level I/II): 2.X1
=
K1 X4=
M+
2K1+
2K2X2
=
K2 X5=
M+
K1+
2K2X3
=
M+
K1+
K2 X6=
M+
2K1+
K2Case Study: Level I Vs. Level II
S T R
X1
X2
X3
X4
X5
X6
A1
= {
1,
4}
, A2= {
2,
3,
4}
, A3= {
1,
2,
5,
6}
, and A4= {
3,
5,
6}
. Routing bound=
3. Shannon bounds: 2 (Level-I) and 3 (Level-II).X1
=
M+
K1 X4=
K1+
K2X2
=
K2 X5=
M+
K1+
K2X3
=
K1Case Study: Level I Vs. Level II
S T R
X1
X2
X3
X4
X5
X6
A1
= {
1,
4}
, A2= {
2,
3,
4}
, A3= {
1,
2,
5,
6}
, and A4= {
3,
5,
6}
. Routing bound=
3. Shannon bounds: 2 (Level-I) and 3 (Level-II).X1
=
M+
K1 X4=
K1+
K2X2
=
K2 X5=
M+
K1+
K2X3
=
K1Heuristic Observations
Objective:H(K) ≥3H(M) Constraints from network:
I
(
M;
K) =
0H
(
X1,
X2,
X3|
M,
K) =
0 H(
X4,
X5,
X6|
X1,
X2,
X3) =
0 H(
M,
K|
X4,
X5,
X6) =
0 I(
M;
X2,
X4,
X5) =
0 I(
M;
X2,
X3,
X6) =
0 I(
M;
X1,
X5,
X6) =
0 I(
M;
X1,
X3,
X4) =
0 I(
M;
X1,
X2,
X4,
X6) =
0More useful constraints:
H
(
X1,
X2,
X3) =
H(
X1) +
H(
X2) +
H(
X3)
H(
X4,
X5,
X6) =
H(
X4) +
H(
X5) +
H(
X6)
H(
X4|
X1,
X3) =
0H
(
X5|
X1,
X2,
X3) =
0 H(
X6|
X2,
X3) =
0 H(
X3) =
2H(
X1)
H(
X1) =
H(
X2)
H(
X5) =
2H(
X4)
H(
X4) =
H(
X6)
Helpful in designing a LNC!
Heuristic Observations (cont’d)
Initial constraints:
H(K) =3H(M) I
(
M;
K) =
0H
(
X1,
X2,
X3|
M,
K) =
0 H(
X4,
X5,
X6|
X1,
X2,
X3) =
0 H(
M,
K|
X4,
X5,
X6) =
0 I(
M;
X2,
X4,
X5) =
0 I(
M;
X2,
X3,
X6) =
0 I(
M;
X1,
X5,
X6) =
0 I(
M;
X1,
X3,
X4) =
0 I(
M;
X1,
X2,
X4,
X6) =
0⇒
Implications:
H
(
X1,
X2,
X3) =
H(
X1) +
H(
X2) +
H(
X3)
H(
X4,
X5,
X6) =
H(
X4) +
H(
X5) +
H(
X6)
H(
X4|
X1,
X3) =
0H
(
X5|
X1,
X2,
X3) =
0 H(
X6|
X2,
X3) =
0 H(
X3) =
2H(
X1)
H(
X1) =
H(
X2)
H(
X5) =
2H(
X4)
H(
X4) =
H(
X6)
Necessary conditions!
A Hard Example: |A| =
12S T R
X1
X2 X3
X4
X5 X6
A1
= {
3,
5,
6}
, A2= {
3,
4,
6}
, A3= {
3,
4,
5}
, A4= {
2,
5,
6}
, A5= {
2,
4,
6}
, A6= {
2,
3,
6}
, A7= {
2,
3,
5}
, A8= {
2,
3,
4}
, A9= {
1,
5,
6}
, A10= {
1,
3,
5}
, A11= {
1,
3,
4}
, A12= {
1,
2,
4,
5}
Routing bound: 4. Shannon bound: 19/5. LNC onF
24q .H
(
X1,
X2,
X3) =
H(
X1) +
H(
X2) +
H(
X3)
H(
X4,
X5,
X6) =
H(
X4) +
H(
X5) +
H(
X6)
7H(
X1) =
9H(
X2)
8H
(
X1) =
9H(
X3) ⇒
9:
7:
8 3H(
X4) =
4H(
X5)
5H
(
X4) =
4H(
X6) ⇒
8:
6:
10Conclusion
F. Cheng and V. Y. F. Tan, “A Numerical Study on the Wiretap Network with a Simple Network Topology,” arxiv:1505.02862.
• A simple network model
• Numerical study by ITIP/Xitip
• Initial progress:
|A| =
4• Some very interesting wiretap patterns
• Open