1: \documentclass[twocolumn,showpacs,amsmath,amssymb]{revtex4}
2: \hoffset -0.5cm \topmargin -35pt
3: \usepackage{graphicx}
4: %============ Text ===============
5:
6: \begin{document}
7:
8: %\preprint{}
9:
10: %\baselineskip 30pt
11:
12: \title{\bf A repair strategy to the attacked random and scale-free networks}
13:
14: \author{L.P. Chi$^\dag$, C.B. Yang and X. Cai}
15: \affiliation{Institute of Particle Physics, HuaZhong Normal
16: University, Wuhan 430079, China \\
17: $^\dag$Email: chilp@iopp.ccnu.edu.cn\\
18: %$^\dag$Tel: 027-62513912\\
19: }
20:
21: \date{\today}
22:
23: \begin{abstract}
24:
25: With a simple attack and repair evolution model, we investigate
26: the stability and structural changes of the Erd\"{o}s-R\'{e}nyi
27: random graphs (RG) and Barab\'{a}si-Albert scale-free (SF)
28: networks. We introduce a new quantity, \textit{invulnerability}
29: $I(s)$, to describe the stability of the system. We find that both
30: RG and SF networks can evolve to a stationary state. The
31: stationary value $I_{c}$ has a power-law dependence on the repair
32: probability $p_{re}$. We also analyze the effects of the repair
33: strategy to the attack tolerance of the networks. We observe that
34: there is a threshold, $(k_{max})_c$, for the maximum degree. The
35: maximum degree $k_{max}$ at time $s$ will be no smaller than
36: $(k_{max})_c$. We give further information on the evolution of the
37: networks by comparing the changes of the topological parameters,
38: such as degree distribution $P(k)$, average degree $\langle k
39: \rangle$, shortest path length $L$, clustering coefficient $C$,
40: assortativity $r$, under the initial and stationary states.
41:
42: \end{abstract}
43:
44: \pacs{89.75.Hc, 87.23.Kg, 89.75.Fb}
45: %89.75.Hc Networks and genealogical trees
46: %87.23.Kg Dynamics of evolution
47: %89.75.Fb Structures and organization in complex systems
48:
49: \maketitle
50:
51: \section{Introduction}
52:
53: Most of the systems in the real world can be represented as
54: complex networks \cite{s1, s2, s3, s4, s5, s6, s7,s8}, in which
55: the nodes are the elementary components of the system and the
56: edges connect pair of nodes that mutually interact exchanging
57: information. The existing empirical and theoretical results
58: indicate that complex networks can be divided into two major
59: classes based on their degree distribution $P(k)$ \cite{s10}. The
60: degree of a node is the number of links adjacent to the node and
61: the degree distribution gives the probability that a node in the
62: network is connected to $k$ other nodes. The first class of
63: networks is characterized by Poisson degree distribution, such as
64: Erd\H{o}s and R\'{e}nyi random graph \cite{s11}. It is constructed
65: starting from an initial condition of $N$ nodes and then adding
66: links with connection probability $p$. The second class of
67: networks is characterized by power-law degree distribution $P(k)
68: \sim k^{-\gamma}$, such as Barab\'{a}si and Albert scale-free
69: network \cite{s12}. It starts from an initial condition of $m_0$
70: nodes and then, for each time step, evolves adding a new node with
71: $m$ edges (growth) that is connected more likely to nodes with
72: higher degree (preferential attachment).
73:
74: Recently enormous interest has been devoted to study the
75: robustness of complex networks under random failures or
76: intentional attacks \cite{s13,s14,s15,s16,s17,s18}. By random
77: failure we mean the removal of randomly selected network nodes and
78: their edges. Instead, we call intentional attack the targeted
79: removal of the most important nodes and their edges. The
80: importance of a node can be determined by the degree, or the
81: betweenness (the number of shortest paths that pass through the
82: node), \textit{etc}.
83:
84: In this paper, the node with highest degree suffers attack. Since
85: this node has highest connections to other nodes, it is most
86: likely attacked and is the most vulnerable site in the network.
87: The aim of this paper is to investigate and compare the stability
88: of Erd\H{o}s and R\'{e}nyi random graph and Barab\'{a}si and
89: Albert scale-free network under the attack and repair strategy. At
90: the same time, we will compare the effects of the attack and
91: repair strategy to these two different topological structures. The
92: paper is organized as follows. In Section II, we describe the
93: evolution model of the attack and repair strategy. Simulation
94: results are in the next section. In the conclusion, we formulate
95: the main results of the paper.
96:
97: \section{The attack and repair strategy}
98:
99: Our attack and repair strategy is simulated as follows. We start
100: by constructing a network according to the algorithms of Erd\H{o}s
101: and R\'{e}nyi random graph (RG) or Barab\'{a}si and Albert
102: scale-free network (SF). In RG, the network size is $N$ and the
103: connection probability is $p$. In SF networks, the initial $m_0$
104: and $m$ are both 2. The dynamics of our model is defined in terms
105: of the following two operations.
106:
107: \begin{enumerate}
108:
109: \item[(i)] \textit{Attack}: Find a node with the
110: maximum degree $k_{max}$ and remove all
111: its links. (If several nodes happen to have the same highest
112: degree of connection, we randomly choose one of them.)
113:
114: \item[(ii)] \textit{Repair}: Reconnect this node
115: with the other nodes in the network with probability $p_{re}$. We
116: assume that all information on the former links of the damaged
117: nodes has been lost, the damaged nodes have be connected randomly
118: to other nodes in the network again.
119:
120: \end{enumerate}
121:
122: Then, the evolution comes into the next Monte Carlo time step. An
123: issue should be mentioned here. Unlike previous attack model
124: \cite{s12} the nodes damaged will not be removed from the network,
125: instead they will be reconnected in certain way. In this way, we
126: keep the size of the system constant.
127:
128: \section{simulation results}
129:
130: \begin{figure}
131: \begin{center}
132: \includegraphics[width=7cm]{fig1.eps}
133: \caption{Plots of $k_{max}$ versus time step $s$ with size
134: $N=10000$ and $p_{re}=0.005$ for (a) random graphs; (b) scale-free
135: networks.}
136: \end{center}
137: \label{fig1}
138: \end{figure}
139:
140: We have performed extensive numerical simulations to study the
141: evolution of RG and SF networks under the attack and repair
142: strategy. In Fig. 1, we give a snapshot of the maximum degree
143: $k_{max}$ versus time step $s$ with $N=10000$ nodes and repair
144: probability $p_{re}=0.005$. We find that $k_{max}$ of RG model
145: decreases slowly with big fluctuations during the evolution. For
146: SF model, $k_{max}$ decreases steeply at the beginning of the
147: evolution and then slowly later. This behavior is rooted in their
148: different topological structures. The RG network is a homogenous
149: network and all nodes have approximately the same number of links.
150: Thus the removal of each node causes the same amount of damage.
151: The SF networks is an extremely inhomogeneous network, in which
152: the minority of nodes have most of the links. The removal of these
153: 'important' nodes will have dramatic impact on the changes of
154: $k_{max}$.
155:
156: Generally, the systems with larger $k_{max}$ have higher
157: probability to be attacked. In order to evaluate the probability
158: that the system is vulnerable, we introduce the concept,
159: \textit{invulnerability}, $I(s)$. Considering an evolution of
160: network with maximum degrees $k_{max}(1)$, $k_{max}(2)$,...,
161: $k_{max}(s)$, invulnerability $I(s)$ at time $s$ is defined as
162:
163: \begin{equation}
164: I(s)=1/Min\{k_{\rm max}(i)\} \ \ \textbf{for\ \ \ } i\leq s\ ,
165: \label{invul}
166: \end{equation}
167:
168: \noindent According to Eq.~(\ref{invul}), $I(s)$ is a
169: non-decreasing parameter, and helps filter out some fluctuations
170: in $k_{max}$.
171:
172: \begin{figure}
173: \begin{center}
174: \includegraphics[width=7cm]{fig2.eps}
175: \caption{Plots of $I(s)$ versus step $s$ with $N=10000$ and
176: $p_{re}=0.005$ for (a) random graphs; (b) scale-free networks.}
177: \end{center}
178: \label{fig2}
179: \end{figure}
180:
181: In Fig. 2 we plot the evolution of $I(s)$ versus $s$ for RG and SF
182: networks under $N=10000$ and $p_{re}=0.005$. We observe that in
183: both networks $I(s)$ increases very quickly at small $s$ but
184: slowly at large $s$ and finally reaches a stationary value $I_{c}$
185: after taking a long transient period. The stationary value $I_c$
186: of the RG network is 0.037, while $I_c$ of the SF network is
187: 0.042.
188:
189: Additionally, we can see from Eq.~(\ref{invul}) that the maximum
190: degree $k_{max}$ of the network in the stationary state is
191:
192: \begin{equation}
193: (k_{max})_{c}=\frac{1}{I_c}.
194: \label{kc}
195: \end{equation}
196:
197: \noindent Since stationary value $I_c$ is the maximum of all the
198: $I(s)$, $k_{max}$ at time $s$ will be no smaller than
199: $(k_{max})_c$, that is, $k_{max}(s)\geq (k_{max})_c$. For
200: instance, under $N=10000$ and $p_{re}=0.005$, the repair strategy
201: ensures the maximum degree $k_{max}$ of the RG network is no
202: smaller than $1/I_c=1/0.037=27$ and the $k_{max}$ of the SF
203: network is no smaller than $1/I_c=1/0.042=24$. The introduction of
204: the repair strategy ensures that $k_{max}$ of the attacked RG and
205: SF networks are greater than 27 and 24 during the evolution,
206: respectively. If without the repair strategy, all the links will
207: be removed after about 5000 time steps for attacked RG model, 2000
208: time steps for attacked SF model.
209:
210: \begin{figure}
211: \begin{center}
212: \includegraphics[width=6cm]{fig3.eps}
213: \caption{Plot of $I_{c}$ versus repair probability $p_{re}$ with
214: network size $N=2000$ (triangles); 3000 (squares); and 5000
215: (circles) for RG and SF networks.}
216: \end{center}
217: \label{fig3}
218: \end{figure}
219:
220: In order to understand the effects of the repair strategy to the
221: stationary state of the system, we plot the stationary value $I_c$
222: as a function of the repair probability $p_{re}$ in log-log scale
223: for RG networks (open symbols) and SF networks (filled symbols)
224: with different network sizes. Triangles, diamonds and circles
225: represent $N=2000$, $3000$ and $5000$, respectively. We find that
226: under fixed network size $N$, the higher the repair probability
227: $p_{re}$ is, the smaller the stationary value $I_c$ is. On the
228: other hand, under the fixed $p_{re}$, the larger $N$ is, the
229: smaller $I_c$ is. In fact, $I_c$ has an inverse ratio to the
230: product of $N$ and $p_{re}$.
231:
232: Figure 2 also tells that the stationary value $I_c$ has a
233: power-law dependence on the repair probability $p_{re}$,
234:
235: \begin{equation}
236: I_c \sim p_{re}^{-\tau}.
237: \end{equation}
238:
239: \noindent Under the same network size $N$ and repair probability
240: $p_{re}$, the stationary values $I_c$ of RG and SF networks are
241: very similar, with $I_c$ of SF networks a little larger, thus more
242: invulnerable.
243:
244: By investigating the changes of the topological parameters of RG
245: and SF networks at the initial and stationary states, we are
246: trying to understand the effects of the repair strategy to the
247: topological structures of these two different structural networks.
248:
249: \textit{Degree distribution}. Figure 4(a) and 4(b) display the
250: degree distributions $P(k)$ of RG and SF networks in the initial
251: and stationary states under $N=5000$ and $p_{re}=0.02$. Figure
252: 4(a) and 4(b) show that the shape of the degree distributions is
253: almost the same for RG and SF networks in initial and stationary
254: states. For RG networks, the degree distributions in the initial
255: and stationary state are both Gaussian distributions, with the
256: peak changes from 100 in the initial state to 38 in the stationary
257: one. At the same time, the distribution of SF networks in the
258: initial and stationary states are both power-law distribution,
259: having the same exponent about 2.6. Figure 4 suggests that the
260: general structures of the systems do not change even after a long
261: time evolution to the stationary states.
262:
263: \begin{figure}
264: \begin{center}
265: \includegraphics[width=7cm]{fig4.eps}
266: \caption{A comparison of degree distributions under $N=5000$ and
267: $p_{re}=0.02$ in initial (circles) and stationary (pluses) states
268: for (a) random graphs; and (b) scale-free networks. Solid lines
269: are the fit curves. }
270: \end{center}
271: \label{fig4}
272: \end{figure}
273:
274: \textit{Average degree}. We find that the average degree of RG
275: network in the stationary state $\langle k \rangle$ decreases
276: dramatically, while slightly for SF networks. This result is
277: caused by the homogeneity of RG networks. Since in RG networks all
278: nodes have approximately the same number of links, majority of
279: nodes in the network will be impacted by the attacks and, hence,
280: the rapid decrease of $\langle k \rangle$. The power-law topology
281: of SF networks displays that $\langle k \rangle$ is mainly
282: determined by the nodes with small degrees, which are less
283: influenced by attacks.
284:
285: \textit{Shortest path length}. The shortest path length $L$ is the
286: average length of the shortest paths between any two nodes in the
287: network. $L$ in the stationary states for RG and SF networks both
288: increases. That is, the efficiency for the nodes in the network to
289: communicate decreases under the attacks.
290:
291: \textit{Clustering coefficient}. The clustering coefficient of
292: node $i$ is defined as the existing numbers $n_{i}$ among the
293: links of node $i$ over all the possible links, that is,
294: $C_{i}=n_{i}/[k_{i}(k_{i}-1)/2]$. The clustering coefficient of
295: the whole network is the average over all the $C_i$. The increase
296: of clustering coefficient for RG networks in the stationary states
297: means the nodes have the tendency to be highly clustered. In fact,
298: the small shortest path length and large clustering coefficient
299: suggest the RG network in the stationary states is more or less a
300: small-world network. The clustering coefficient of SF networks
301: decreases a little in the stationary states. It seems that few
302: nodes with largest connections are attacked and have a big affect
303: on the network, which makes the connections in the network becomes
304: sparse, thus the clustering coefficient becomes decrease.
305:
306: \textit{Assortativity}. The assortativity $r$ in the range $-1
307: \leq r \leq 1$ is another interesting feature of complex networks.
308: A network is said to show assortative, $r>0$, if the high-degree
309: nodes in the network preferentially connect to other high-degree
310: nodes. By contrast, a network is disassortative, $r<0$, if the
311: high-degree nodes tend to connect to other low-degree nodes. The
312: increase of $r$ for RG networks in the stationary states are
313: consistent with our conjecture that the networks becomes a little
314: highly clustered. For SF networks, they are inhomogeneous in both
315: initial and stationary states according to the analysis of the
316: degree distribution above, so $r$ are both negative, that is
317: disassortative.
318:
319: \begin{table}[t]
320: \caption{\em Comparison of topological changes in initial and
321: stationary states between RG and SF networks under $N=5000$ and
322: $p_{re}=0.02$.}
323: \begin{center}
324: \begin{tabular}{c|c|c|p{0.15cm}|c|c}
325: \hline\hline
326: &\multicolumn{2}{c|}{\bf{RG}} & & \multicolumn{2}{|c}{\bf{SF}} \\
327: \cline{2-3}\cline{5-6}
328: & initial & stationary & & initial & stationary\\
329: \hline
330: $I_{c}$ & --- & 0.0149 & & --- & 0.0152 \\
331: $\langle k \rangle$ & 100.06 & 38.09 & & 3.99 & 3.92 \\
332: $ L $ & 2.11 & 3.89 & & 4.83 & 5.28\\
333: $ C $ & 0.02 & 0.47 & & 0.0068 & 0.0028\\
334: $ r $ & 0.23 & 0.52 & & -0.062 & -0.074\\
335: \hline
336:
337: \end{tabular}
338: \end{center}
339: \label{comparison}
340: \end{table}
341:
342: \section{Conclusions}
343:
344: Summarily, in this paper we study the effects of the repair
345: strategy to the attacked ER random graph model and BA scale-free
346: model. We find the introduction of the repair strategy has
347: significant effects on the attack tolerance of the networks. We
348: also introduce a new quality, \textit{invulnerability} $I(s)$, to
349: describe the stability of the networks. After a long time
350: evolution, the invulnerability of ER random graph model and BA
351: scale-free model can both evolve to a stationary value $I_c$.
352: $I_c$ has a power-law dependence on the repair probability
353: $p_{re}$. In addition, the maximum degree in the stationary state
354: $(k_{max})_c$ is always smaller than the maximum degree at time
355: $s$, $k_{max}(s)$.
356:
357: We give further information on the evolution of the networks by
358: comparing the changes of the topological parameters, such as
359: degree distribution $P(k)$, average degree $\langle k \rangle$,
360: shortest path length $L$, clustering coefficient $C$ and
361: assortativity $r$, between the initial and stationary states. We
362: conclude that the topological structure of SF networks does not
363: change during the evolution, while RG networks become small-world
364: networks during the evolution, according to the small shortest
365: path length and large clustering coefficient.
366:
367: \section*{ACKNOWLEDGEMENTS}
368:
369: This work was supported by National Natural Science Foundation of
370: China under Grant Nos. 70401020, 70571027 and 10647125.
371:
372:
373: \begin{thebibliography}{s1}
374:
375: \bibitem{s1}
376: S. Boccaletti, V. Latora, Y. Moreno, \textit{et al},
377: \textit{Physics Reports} \textbf{424} (2006) 175.
378:
379: \bibitem{s2}
380: R.J. Williams and N.D. Martinez, \textit{Nature} \textbf{404}
381: (2000) 180.
382:
383: \bibitem{s3}
384: W. Li and X. Cai X, \textit{Phys. Rev.} E \textbf{69} (2004)
385: 046106.
386:
387: \bibitem{s4}
388: L.P. Chi, R. Wang, H. Su, \textit{et al}, \textit{Chin. Phys.
389: Lett.} \textbf{20} (2003) 1393.
390:
391:
392: \bibitem{s5}
393: L.P. Chi and X. Cai, \textit{Int. J. Mod. Phys.} B \textbf{18}
394: (2004) 2394.
395:
396: \bibitem{s6}
397: V. Latora and M. Marchiori, \textit{Physica} A \textbf{314} (2002)
398: 109.
399:
400: \bibitem{s7}
401: R.F. Cancho and R.V. Sol\'{e}, \textit{Proc. Natl. Acad. Sci.}
402: \textbf{100} (2003) 788.
403:
404: \bibitem{s8}
405: Xiang Li and G. Chen, \textit{Physica A} \textbf{328} (2003) 274.
406:
407: \bibitem{s9}
408: R. Albert, H. Jeong and A.-L. Barab\'{a}si, \textit{Nature}
409: \textbf{401} (1999) 130.
410:
411: \bibitem{s10}
412: A.-L. Barab\'{a}si and R. Albert, Science \textbf{286} (1999) 509.
413:
414: \bibitem{s11}
415: P. Erd\H{o}s and A. R\'{e}nyi, \textit{Publ. Math. Inst. Hung.
416: Acad. Sci.} \textbf{5} (1960) 17.
417:
418: \bibitem{s12}
419: R. Albert, H. Jeong and A.-L. Barab\'{a}si, \textit{Nature}
420: \textbf{406} (2000) 378.
421:
422: \bibitem{s13}
423: R. Albert and A.-L. Barab\'{a}si, \textit{Rev. Mod. Phys.}
424: \textbf{74} (2002) 47.
425:
426: \bibitem{s14}
427: M.E.J. Newman, \textit{Eur. Phys. J.} B \textbf{38} (2004) 321.
428:
429: \bibitem{s15}
430: L.P. Chi, C.B. Yang and X. Cai, \textit{Chin. Phys. Lett.}
431: \textbf{23} (2006) 263.
432:
433: \bibitem{s16}
434: L.P. Chi, C.B. Yang, K. Ma and X. Cai, \textit{Commun. Theor.
435: Phys.} \textbf{45} (2006) 765.
436:
437: \bibitem{s17}
438: P. Crucitti, V. Latora, M. Marchiori and A. Rapisarda, Physica A
439: \textbf{340} (2004) 388.
440:
441: \bibitem{s18}
442: P. Crucitti, V. Latora, M. Marchiori and A. Rapisarda, Physica A
443: \textbf{320} (2003) 622.
444:
445:
446: \end{thebibliography}
447:
448:
449: \end{document}
450: