Presentation is loading. Please wait.

Presentation is loading. Please wait.

Xiaowei Ying, Kai Pan, Xintao Wu, Ling Guo Univ. of North Carolina at Charlotte SNA-KDD June 28, 2009, Paris, France Comparisons of Randomization and K-degree.

Similar presentations


Presentation on theme: "Xiaowei Ying, Kai Pan, Xintao Wu, Ling Guo Univ. of North Carolina at Charlotte SNA-KDD June 28, 2009, Paris, France Comparisons of Randomization and K-degree."— Presentation transcript:

1 Xiaowei Ying, Kai Pan, Xintao Wu, Ling Guo Univ. of North Carolina at Charlotte SNA-KDD June 28, 2009, Paris, France Comparisons of Randomization and K-degree Anonymization Schemes for Privacy Preserving Social Network Publishing

2 SNA-KDD09, June 28, Paris, France Motivation Privacy Preserving Social Network Publishing node-anonymization cannot guarantee identity/link privacy due to subgraph queries. Backstrom et al. WWW07, Hay et al. UMass TR07 edge randomization Random Add/Del, Random Switch K-anonymity generalization Hay et al. VLDB08, K-degree Liu&Terzi SIGMOD08, Zhou&Pei ICDE08 Utility preserving randomization Spectral feature preserving Ying&Wu SDM08 Real space feature preserving based on Markov Chain Ying&Wu SDM09, Hanhijarvi et al. SDM09 2

3 SNA-KDD09, June 28, Paris, France Motivation Attacks based on Background Knowledge Attributes of vertices Vertex degrees Specific link relationships between target individuals Neighborhoods of target individuals Embedded subgraphs Graph metric 3

4 SNA-KDD09, June 28, Paris, France Focus We quantify identity disclosure and link disclosure under vertex degrees attacks for Rand Add/Del. Identity disclosure is measured as the prob. of correctly linking a target individual to an anonymized node given the degree of the target individual. Link disclosure as the prob. of existence of a sensitive link between two individuals given their known degrees. Details skipped We compare Rand Add/Del with K-degree generalization in terms of utility preservation (under the same privacy disclosure threshold, i.e., 1/K) 4

5 SNA-KDD09, June 28, Paris, France Network of US political books (105 nodes, 441 edges) Books about US politics sold by Amazon.com. Edges represent frequent co-purchasing of books by the same buyers. Nodes have been given colors of blue, white, or red to indicate whether they are "liberal", "neutral", or "conservative". http://www-personal.umich.edu/˜mejn/netdata/ 5 Political books network

6 SNA-KDD09, June 28, Paris, France Degree variation due to randomization 6

7 SNA-KDD09, June 28, Paris, France Applying Bayesian Theorem Re-identification risks 7 The attacker does not know the original degree distribution.

8 SNA-KDD09, June 28, Paris, France Estimate original degree sequence 8 Original degree sequenceAfter randomizationEstimated Add & delete 10% edges

9 SNA-KDD09, June 28, Paris, France Node re-identification risks Nodes’ prior and posterior risks Given an individual α with degree d α and a randomized graph Prior risk: Posterior risks 9

10 SNA-KDD09, June 28, Paris, France Re-identification risks Re-identification risks reduces as k increases; Add/Del strategy can efficiently reduce the risk. 10

11 SNA-KDD09, June 28, Paris, France Protection vs. randomization k Node’s absolute and relative protection measures Absolute measure Relative measure 11

12 SNA-KDD09, June 28, Paris, France Comparison K-degree generalization (Liu&Terzi SIGMOD08) to construct a K-degree anonymous graph where every node has the same degree with at least K-1 other nodes. Random Add/Del Determine perturbation magnitude k to satisfy identity disclosure < 1/K, and then perturb graph using k. 12

13 SNA-KDD09, June 28, Paris, France Utility features 13 Largest eigenvalue of Adjacency matrix: λ 1 Second smallest eigenvalue of Laplacian matrix: μ 2 Harmonic mean of shortest distance: Modularity (community structure) Transitivity(cluster coefficient) Subgraph centrality

14 SNA-KDD09, June 28, Paris, France 14

15 SNA-KDD09, June 28, Paris, France Observation Both Rand Add/Del and K-degree generalization decrease structural properties. K-degree generally better preserves structural features K-degree chooses a subset of nodes ( which violate K-degree anonymity) for edge modification while Rand Add/Del treats all nodes/edges equally for randomization We can improve Rand Add/Del by dividing the graph into blocks and apply randomization on each block. (next slide) We expect Rand Add/Del is more robust to other attacks. (ongoing work) We expect reconstruction methods can be designed on the purely randomized graph to recover features accurately. (ongoing work) 15

16 SNA-KDD09, June 28, Paris, France Block Add/Del 16

17 SNA-KDD09, June 28, Paris, France Conclusion Quantify how well Rand Add/Del can protect node identity and link privacy under the vertex degree background knowledge attack Compare with K-degree generalization scheme in terms of utility preservation Future Work Other background knowledge attacks Other randomization schemes Reconstruction methods on the randomized graph 17

18 Questions? Acknowledgments This work was supported in part by U.S. National Science Foundation IIS-0546027 and CNS-0831204. Thank You! 18


Download ppt "Xiaowei Ying, Kai Pan, Xintao Wu, Ling Guo Univ. of North Carolina at Charlotte SNA-KDD June 28, 2009, Paris, France Comparisons of Randomization and K-degree."

Similar presentations


Ads by Google