Vous êtes sur la page 1sur 18

Distributed Consensus in Faulty Networks

Halil Yiğit Öksüz, Mehmet Akar∗

Networked and Embedded Control Systems Laboratory


Department of Electrical and Electronics Engineering
Boğaziçi University
Istanbul, Turkey

(yigit.oksuz,mehmet.akar@boun.edu.tr)

December 13, 2017

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 1 / 18


Overview

1 Linear Consensus Algorithm

2 Graph Representation

3 Fault Tolerant Consensus


Expanded Network Representation
Consensus Model in the Presence of StrBYZ Agents

4 Expanded Network Analysis

5 Original Network Analysis

6 Conclusion

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 2 / 18


Linear Consensus Algorithm

Consider a network of n agents. Each node in the network uses the update
rule X
xi [k + 1] = wij [k]xj [k] (1)
j∈Ni

where wij [k] are averaging coefficients which are assumed to satisfy the
following conditions:
wij [k] = 0 whenever j ∈
/ Ni , i ∈ V , k ∈ Z≥0 ;
0 < α ≤ wij [k], ∀j ∈ Ni , i ∈ V , k ∈ Z≥0 for some parameter α;
Pn
j=1 ωij [k] = 1, ∀i ∈ V , k ∈ Z≥0 .

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 3 / 18


Graph Representation

Directed graph: G = (V , E ) where

? V : node set indexed by the set I = {1, 2, · · · , n}.


? E ⊂ V × V : directed edge set.
? Vnf = V \ F : set of non-faulty nodes in this network where F is the
set of faulty nodes.
? The cardinality of a set Vnf is denoted by |Vnf |.
? Directed edge (i, j) ∈ E : represents the information flow j → i

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 4 / 18


Graph Representation

? xi [k]: state value of node i.


? xij [k] value of j received by node i.
? Ni+ = {j ∈ V : (i, j) ∈ E }: set of in-neighbors of node i.
? Ni− = {j ∈ V : (j, i) ∈ E }: set of out-neighbors of node i.
? Ni = Ni+ ∪ {i}: inclusive neighbors of node i (each node knows its
own state).

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 5 / 18


Primary and Secondary Layer Concepts
Definition 1 (Primary layer [1, 2])
Let G = (V , E ) be a digraph. There exist l1 (l1 ≥ 1) subsets in the vertex set V
such that each subset V1,s , s = 1, ..., l1 , is the largest possible subset that has a
spanning tree for its subgraph G1,s and for all i ∈ V1,s and j ∈ / V1,s , we have
(i, j) ∈
/ E . We say G1,s are the primary layer subgraphs of G where the number of
primary layer subgraphs is denoted by l1 .

Definition 2 (Secondary layer [1, 2])


Let V̄ be the set which consists of the vertices that are not in the primary layer
subgraphs, i.e., V̄ = V \ ls=1
S1
V1,s . Then, there exists l2 subsets in V̄ such that
each subset V2,s , s = 1, ..., l2 , has a spanning tree for its subgraph and there exist
exactly one node i ∈ V2,s which satisfies the following:
For all j ∈ V2,s \ i and m ∈ V \ V2,s , we have (j, m) ∈
/ E.
There exist at least two vertices in two different subgraphs (either primary or
secondary layer) p and v such that (i, p) ∈ E and (i, v ) ∈ E .
i is the root of a spanning tree in V2,s .
We define the subsets V2,s , s = 1, ..., l2 as the secondary layer subgraphs of G .

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 6 / 18


Fault Tolerant Consensus

Definition 3 (Structured Byzantine)


A node i ∈ F is said to be StrBYZ if it sends different information to its
out-neighbors while employing |Ni− | different update rule (1) starting with
|Ni− | different initial conditions.

In fault-tolerant distributed consensus, the objective is to make sure that


all non-faulty agents achieve consensus on a single value in the presence of
misbehaving agents.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 7 / 18


Expanded Network Representation
Since xji [k] 6= xmi [k], i ∈ Nj ∩ Nm , j, m ∈ V holds for all i ∈ F , faulty
agents can be modelled by making appropriate changes on the network
topology as follows:
All existing connections between non-faulty nodes are preserved.
Each faulty node creates one copy of itself for each of its
out-neighbors by preserving its in-neighbours and these copies are
disconnected if they are generated from the same faulty node.

1 1’ 1”

2 3 2 3

Figure: Original and Expanded Networks where node 1 is StrBYZ.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 8 / 18


Consensus Model in the Presence of StrBYZ

We can represent the update rules for both the non-faulty i-th node and
the faulty node j as

X X
xi [k + 1] = wij xj [k] + wij xij [k], i ∈ Vnf (2)
j∈Ni ∩Vnf j∈Ni ∩F

X X
xij [k + 1] = w̃ij,l xl [k] + w̃ij,l xjl [k], j ∈ F (3)
l∈Nj ∩Vnf l∈Nj ∩F

where xij [k] represents the information sent by the faulty node j to the
non-faulty node i at the time step k. Note that wij and w̃ij,l are the
coefficients that satisfy the same conditions with the coefficients in the
update rule (1).

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 9 / 18


Expanded Network Analysis

Theorem 1
Non-faulty nodes of the network G achieve consensus in the presence of
faulty nodes if and only if they are all in the same primary or secondary
layer in the expanded network Ḡ .

Note: Theorem 1 states a condition that must be satisfied in the


expanded network Ḡ in order for non-faulty nodes to achieve consensus in
the original network G .

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 10 / 18


Original Network Analysis

Lemma 1
Non-faulty nodes of the network G are all in the same primary layer in the
expanded network Ḡ if and only if the following conditions hold:
(i) Non-faulty nodes are all in the same primary layer (V1,s ) in the
original network G , i.e., Vnf ⊆ V1,s where s ∈ {1, 2, · · · , l1 }, l1 ≥ 1.
(ii) Each faulty node in this primary layer V1,s has at least one
in-neighbor, i.e., for each i ∈ V1,s ∩ F , |Ni+ | ≥ 1.

Note: Lemma 1 states conditions on the original network G for all


non-faulty nodes to be in the same primary layer in the expanded network
Ḡ .

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 11 / 18


Original Network Analysis

Lemma 2
Non-faulty nodes of the network G are all in the same secondary layer in
the expanded network Ḡ if and only if the following conditions hold:
(i) Non-faulty nodes are all in the same secondary layer V2,s having a
non-faulty root node in the network G , i.e., Vnf ⊆ V2,s where
s ∈ {1, 2, · · · , l2 }, l2 ≥ 0.
(ii) Each faulty node in the secondary layer V2,s has at least one
in-neighbor, i.e., for each i ∈ V2,s ∩ F , |Ni+ | ≥ 1.

Note: Lemma 2 states conditions on the original network G for all


non-faulty nodes to be in the same secondary layer in the expanded
network Ḡ .

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 12 / 18


Main Result

Theorem 2
Non-faulty nodes of the network G achieve consensus in the presence of
faulty nodes if and only if the following conditions hold:
(i) Non-faulty nodes are either all in the same primary layer (V1,s ) or in
the same secondary layer (V2,s ) having a non-faulty root node, i.e.,
Vnf ⊆ Vl,s where l ∈ {1, 2}.
(ii) Each faulty node in the same layer that contains all non-faulty nodes
in the network G has at least one in-neighbor, i.e., for each
i ∈ Vl,s ∩ F , |Ni+ | ≥ 1.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 13 / 18


Example 1

2.5
x1 [k]
x2 [k]
x3 [k]
2 x4 [k]
x5 [k]

State Values xi [k]


4’ 1.5

1 2 1 2 1

0.5
0 5 10 15 20 25 30
3 4 3 4” k

(a) Original (b) Expanded (c) Simulation Result


network network

Figure: Original and Expanded Networks where node 4 is StrBYZ.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 14 / 18


Conclusion

X We have presented topological conditions for non-faulty nodes to


achieve consensus in the presence of StrBYZ agents.
X Necessary and sufficient conditions have been presented for primary
and secondary layers to be preserved in the presence of StrBYZ
agents.
X Depending on the network structure, non-faulty nodes of the network
reach consensus without employing any resilient algorithm.
X Hence, nonuse of a resilient algorithm reduces the complexity of the
algorithm and computational power in real world applications.
X Theoretical results have been illustrated by an numerical example.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 15 / 18


Acknowledgement

This work was sponsored by TUBITAK Project 114E613.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 16 / 18


References
Erkan, Ö. Feyza, Onur Cihan, and Mehmet Akar. ”Analysis of Distributed Consensus Protocols with Multi-Equilibria
under Time-Delays.” Journal of the Franklin Institute (2017).
Erkan, Ö. Feyza, Onur Cihan, and Mehmet Akar. ”Distributed consensus with multi-equilibria in directed networks.”
American Control Conference (ACC), 2017.
H. J. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram, “Resilient asymptotic consensus in robust networks,” IEEE
Journal on Selected Areas in Communications, vol. 31, no. 4, pp. 766–781, 2013.
H. J. LeBlanc and X. Koutsoukos, “Resilient asymptotic consensus in asynchronous robust networks,” in 50th Annual
Allerton Conference on Communication, Control, and Computing, 2012, pp. 1742–1749.
A. Haseltalab and M. Akar, “Approximate byzantine consensus in faulty asynchronous networks,” in American Control
Conference, 2015, pp. 1591–1596.
——, “Convergence rate analysis of a fault-tolerant distributed consensus algorithm,” in 54th IEEE Annual Conference
on Decision and Control, 2015, pp. 5111–5116.
F. Pasqualetti, A. Bicchi, and F. Bullo, “Distributed intrusion detection for secure consensus computations,” in 46th
IEEE Conference on Decision and Control, 2007, pp. 5594–5599.
——, “Consensus computation in unreliable networks: A system theoretic approach,” IEEE Transactions on Automatic
Control, vol. 57, no. 1, pp. 90–104, 2012.
H. Zhang and S. Sundaram, “Robustness of information diffusion algorithms to locally bounded adversaries,” in
American Control Conference, 2012, pp. 5855–5861.
A. W. Krings and Z. Ma, “Fault-models in wireless communication: Towards survivable ad hoc networks,” in Military
Communications Conference, 2006, pp. 1–7.
M. H. Azadmanesh and R. M. Kieckhafer, “Exploiting omissive faults in synchronous approximate agreement,” IEEE
Transactions on Computers, vol. 49, no. 10, pp. 1031–1042, 2000.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 17 / 18


References
R. Kieckhafer and M. H. Azadmanesh, “Reaching approximate agreement with mixed-mode faults,” IEEE Transactions
on Parallel and Distributed Systems, vol. 5, no. 1, pp. 53–63, 1994.
L. Lamport, R. Shostak, and M. Pease, “The byzantine generals problem,” ACM Trans. Program. Lang. Syst., vol. 4,
no. 3, pp. 382–401, Jul. 1982.
L. Tseng and N. Vaidya, “Iterative approximate byzantine consensus under a generalized fault model,” in International
Conference on Distributed Computing and Networking. Springer, 2013, pp. 72–86.
——, “Iterative approximate consensus in the presence of byzantine link failures,” in Networked Systems. Springer,
2014, pp. 84–98.
N. H. Vaidya, L. Tseng, and G. Liang, “Iterative approximate byzantine consensus in arbitrary directed graphs,” in
Proceedings of the ACM symposium on Principles of distributed computing, 2012, pp. 365–374.
N. H. Vaidya, “Iterative byzantine vector consensus in incomplete graphs,” in International Conference on Distributed
Computing and Networking. Springer, 2014, pp. 14–28.
L. Ju, H. Li, Y. Liu, W. Xue, K. Li, and Z. Chi, “An improved intrusion detection scheme based on weighted trust
evaluation for wireless sensor networks,” in Proceedings of the 5th International Conference on Ubiquitous Information
Technologies and Applications, 2010, pp. 1–6.
B. Parno, A. Perrig, and V. Gligor, “Distributed detection of node replication attacks in sensor networks,” in IEEE
Symposium on Security and Privacy, 2005, pp. 49–63.
W. Ren and R. W. Beard, “Consensus seeking in multiagent systems under dynamically changing interaction topologies,”
IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655–661, 2005.
T. Wang, Y. Zhang, J. Qiu, and H. Gao, “Adaptive fuzzy backstepping control for a class of nonlinear systems with
sampled and delayed measurements,” IEEE Transactions on Fuzzy Systems, vol. 23, no. 2, pp. 302–312, 2015.
H. Gao, T. Chen, and J. Lam, “A new delay system approach to network-based control,” Automatica, vol. 44, no. 1, pp.
39–52, 2008.

Halil Yiğit Öksüz, Mehmet Akar∗ December 13, 2017 18 / 18

Vous aimerez peut-être aussi