The (not so) Critical Nodes of Criminal Networks Donatella Firmani1 , Giuseppe F. Italiano1 , Luigi Laura2 1 University of Rome “Tor Vergata” 2 Sapienza University of Rome CrimeNet 2014 - Barcellona, EspanĚa - November 10th , 2014 Motivations “Criminal networks are not simply social networks operating in criminal contexts. The covert settings that surround them call for specific interactions and relational features within and beyond the network.” Carlo Morselli, Inside Criminal Networks. Springer, 2009. Peculiar features of Criminal Networks Sparrow, in a seminal paper [23], listed four peculiarities: I limited size; I information incompleteness (i.e., criminal network data is inevitably incomplete); I undefined borders, i.e. it is not easy to discover all the connections of a node; I dynamics, that is, many of the useful networks questions depends heavily on the temporal dimension. In this scenario, SNA tools are limited, due to this intrinsic differences! Who are the key players in a network? I One of the most basic questions associated to the analysis of a network is who are the key players, or who are the central nodes in the network? I Node-centrality measures, such as degree, betweenness and closeness centrality, can successfully identify most key players in social networks (see the survey of Borgatti, [8]). I On the contrary, in criminal networks, the most important actors do not necessarily display high centrality scores (Baker and Faulkner, [4]). Our contribution We show that that criminal networks can suffer high disconnection when few nodes are deleted. However, while in social networks nodes whose removal disconnects most nodes from the network belong to their cores, in criminal networks this is not the case. This suggests that: I nodes that pulls together the network deliberately operate from network peripheries, thus being protected from detection; I the key players in the network may not be the ones whose removal affects large portions of the network. Preliminaries: Graph connectivity D A F B J G C E H M K I N L An example graph with two connected components and four articulation points (B, G, H, K). Preliminaries: Graph Cores An example graph, its 2-core (light and dark grey nodes), and its 3-core (dark grey nodes). Nodes in the 3-core are a subset of the nodes in the 2-core of the graph. Preliminaries: Critical Nodes I The concept of critical nodes has been introduced by Ausiello et al. in [2], building upon the concepts of articulation points and core of a connected network. I Precisely, the critical nodes are defined as the articulation points belonging to the network core. Using different samples of the Autonomous Systems network, Ausiello et al. show that: I I I the removal of few critical nodes can affect large portions of the network, thus they are central in a very strong sense [2]; the critical nodes have orders of magnitude higher centrality scores than other nodes [3]. Relations between Articulation Points and Critical Nodes If we focus on the 2-core, we can state the following Lemma Let G be an undirected connected graph. The critical nodes of G (i.e. the articulation points of G 2 ) are articulation points of G . [The proof is in the paper] Relations between Articulation Points and Critical Nodes If we focus on the 3-core, we can say that the previous relation does not hold, as shown in this figure The red node is not an articulation point of the graph, but it is after the removal of the blue node (that is the only node removed when we consider the 3-core of the graph). The network datasets Name Drug-Traffic Terrorists Karate Science Facebook Lindenstrasse Marvel Airlines PowerGrid ASes Type Criminal Network Criminal Network Social Network Social Network Social Network Fictitious SN Fictitious SN Other Other Other Nodes (n) 2749 4275 34 1589 4039 233 10822 235 4941 22963 Edges (m) 13578 7874 78 2742 88234 325 314054 1297 6594 48436 Source Mainas [15] Mainas [15] GEPHI [13] Pajek [5] SNAP [22] Pajek [5] GEPHI [13] GEPHI [13] GEPHI [13] GEPHI [13] Experimental results We measured, for each dataset: 1. the number of articulation points, in short APs; 2. the number of critical nodes, in short CNs; 3. the number of critical nodes in the 3-core, in short CN3 s; 4. the number of CNs and CN3 s that belong to the top K articulation points, sorted by the impact, where K is respectively the number of CNs or CN3 s. Experimental Results: Articulation Points and Critical Nodes G Drug-Traffic Terrorists Karate Science Facebook Lindenstrasse Marvel Airlines PowerGrid ASes n 2749 4275 34 1589 4039 233 10822 235 4941 22963 m 13578 7874 78 2742 88234 325 314054 1297 6594 48436 n 1554 4085 34 379 4039 233 10822 235 4941 22963 largest CC m 2216 6358 78 914 88234 325 314054 1297 6594 48436 APs 280 521 1 57 11 74 107 9 1229 1870 n 454 1303 33 352 3964 123 10543 201 3353 14966 G2 m 1116 3576 78 887 88159 215 313775 1263 5006 40439 APs 15 77 1 44 7 1 3 2 94 10 n 163 681 22 265 3856 12 9935 162 231 4383 G3 m 588 2440 55 736 87952 19 312565 1190 479 19678 APs 0 7 1 1 1 0 2 1 1 1 Metric: Impact of Articulation Point We define impact of a node v is given by the number of nodes that get disconnected from the largest connected component when v is removed. We use the impact to derive two ratios: |CNs∩(Top APs)| |CNs| I membership ratio: I weighted impact ratio: the sum of the impacts of the CNs (or of the CN3 ) divided by the sum of the impacts of the top K APs. An example I The Marvel network has 107 APs, 3 CN, and 2 CN3 s. I The top APs have impacts: {39, 27, 15, 13, 12, . . .}. I The three CNs have impact values equals to {39, 4, 3}. I Of the two CN3 s, none belongs to the top two APs. Thus: I The membership ratio of the CNs is 1/3. I The weighted membership ratio of the CNs is 39/(39 + 27 + 15). I The membership ratio and weighted membership ratio of the CN3 s is zero. Impact of Articulation Points and Critical Nodes Drug-Traffic Terrorists Karate Science Facebook Lindenstrasse Marvel Airlines PowerGrid ASes lCC APs CNs 280 521 1 57 11 74 107 9 1229 1870 15 77 1 44 7 1 3 2 94 10 G2 CNs in TOP APs 8 40 1 44 7 1 1 1 47 10 CN3 s 0 7 1 1 1 0 2 1 1 1 G3 CN3 s in TOP APs − 2 1 1 1 − 0 1 0 1 Max Impact AP CN CN3 172 83 7 60 197 13 39 11 106 333 57 83 7 60 197 13 39 11 106 333 − 83 7 60 197 − 1 11 8 333 Overall impact of Critical Nodes Criminal Networks Social Networks Fictitious SNs Other 100 80 60 40 20 0 Dr ug -Tr Te aff ic rro ris Ka ts rat e Sc ien ce Fa Lin ce Ma de bo o k ns tr rve as s e l Air lin es Po we AS rG rid es Membership ratio (yellow) and the weighted impact ratio (red). Overall impact of CN3 , the Critical Nodes of the 3-core Criminal Networks Social Networks Fictitious SNs Other 100 80 60 40 20 0 Te rro ris ts Ka rat e Sc ien ce Fa ce bo ok Ma rve l Air lin es Po AS we rG rid es Membership ratio (yellow) and the weighted impact ratio (red). Conclusions We used the notion of Critical Nodes to compare Criminal Networks and Social Networks (and other networks). Our findings are that: I as for the SNs, the removal of few nodes can have a huge impact on the overall network... I ...but, differently from SNs, these nodes do not belong to the core of the network!