secgraph:a uniform and open-source evaluation system for ...secgraph:a uniform and open-source...

SecGraph: A Uniform and Open-source Evaluation System for Graph Data

Anonymization and De-anonymization

Shouling Ji, Weiqing Li, and Raheem Beyah Georgia Ins=tute of Technology

Prateek MiDal Princeton University

Xin Hu IBM Research

S. Ji, W. Li, P. MiDal, X. Hu, and R. Beyah SecGraph: A Uniform and Open-‐Source Evalua=on System 2

Introduc=on •  Graph Data

–  Social networks data

–  Mobility traces

–  Medical data

–  ……

•  Why data sharing/publishing –  Academic research

–  Business applica=ons

–  Government applica=ons

–  Healthcare applica=ons

–  ……


Anonymiza=on and De-‐anonymiza=on

Anonymization

De-anonymization

Utility Anonymization vs DA


Contribu=on •  We studied and analyzed

–  11 state-‐of-‐the-‐art graph anonymiza=on techniques

–  15 modern Structure-‐based De-‐Anonymiza=on (SDA) aDacks

•  We developed and evaluated SecGraph –  11 anonymiza=on techniques

–  12 graph u=lity metrics

–  7 applica=on u=lity metrics

–  15 de-‐anonymiza=on aDacks

•  We released SecGraph –  Publicly available at hDp://www.ece.gatech.edu/cap/secgraph/


Outline

•  Introduc=on •  Graph Anonymiza=on •  Graph De-‐anonymiza=on

•  Anonymiza=on vs. De-‐anonymiza=on Analysis •  SecGraph

•  Conclusion


Graph Anonymiza=on •  Naïve ID Removal

•  Edge Edi=ng (EE) based anonymiza=on –  Add/Del

–  Switch

•  K-‐anonymity –  K-‐Neighborhood Anonymiza=on (k-‐NA)

–  K-‐Degree Anonymiza=on (k-‐DA)

–  K-‐automorphism (k-‐auto)

–  K-‐isomorphism (k-‐iso)

•  Aggrega=on/class/cluster based anonymiza=on

•  Differen=al Privacy (DP) based anonymiza=on

•  Random Walk (RW) based anonymiza=on


U=lity Metrics •  Graph u=lity (12)

–  Degree (Deg.)

–  Joint Degree (JD)

–  Effec=ve Diameter (ED)

–  Path Length (PL)

–  Local Clustering Coefficient (LCC)

–  Global Clustering Coefficient (GCC)

–  Closeness Centrality (CC)

–  Betweenness Central=y (BC)

–  Eigenvector (EV)

–  Network Constraint (NC)

–  Network Resilience

–  Infec=ousness (Infe.)

•  Applica=on u=lity (7) –  Role eXtrac=on (RX)

–  Influence Maximiza=on (IM)

–  Minimum-‐sized Influen=al Node Set (MINS)

–  Community Detec=on (CD)

–  Secure Rou=ng (SR)

–  Sybil Detec=on (SD)

Graph utility captures how the anonymized data preserve fundamental structural

properties of the original graph

How useful the anonymized data are for practical graph applications and mining

tasks


Anonymiza=on vs U=lity Resilient to DA attacks

Anonymization Techniques

Conditionally preserve the utility

Partially preserve the utility

Not preserve the utility

preserve the utility


Anonymiza=on vs U=lity Resilient to DA attacks

Anonymization Techniques

Conditionally preserve the utility

Partially preserve the utility

Not preserve the utility


Outline



•  Conclusion


Graph De-‐Anonymiza=on (DA) •  Seed-‐based DA

–  Backstrom et al.’s aDack (BDK)

–  Narayanan-‐Shma=kov’s aDack (NS)

–  Narayanan et al.’s aDack (NSR)

–  Nilizadeh et al.’s aDack (NKA)

–  Distance Vector (Srivatsa-‐Hicks aDack) (DV)

–  Randomized Spanning Trees (Srivatsa-‐Hicks aDack) (RST)

–  Recursive Subgraph Matching (Srivatsa-‐Hicks aDack) (RSM)

–  Yartseva-‐Grossglauser’s aDack (YG)

–  De-‐Anonymiza=on aDack (Ji et al.’s aDack) (DeA)

–  Adap=ve De-‐Anonymiza=on aDack (Ji et al.’s aDack) (ADA)

–  Korula-‐LaDanzi’s aDack (KL)

•  Seed-‐free DA –  Pedarsani et al.’s aDack (PFG)

–  Ji et al.’s aDack (JLSB)

Anonymized Graph Auxiliary Graph

Seed mappings

Figure source: E. Kazemi et al., Growing a Graph Matching from a Handful of Seeds, VLDB 2015


DA ADacks

SF: seed-free AGF: auxiliary graph-free SemF: semantics-free A/P: active/passive attack Scal.: scalable Prac.: practical Rob.: robust to noise


Outline



•  Conclusion


Anonymiza=on vs. DA

Modern DA Attacks

State-of-the-art Anonymization

Techniques

: vulnerable : conditionally vulnerable : invulnerable

DA attacks!


Outline



•  Conclusion


SecGraph: Overview & Implementa=on

Publishing Data

AnonymizedData

Raw Data

AnonymizationModule (AM)Utility

Module (UM)

DAModule (DM)

15 DA Attacks

19 Utility Metrics

11 Anonymization Techniques

Raw Data -> AM -> Anonymized Data -> UM/DM -> Publishing Data Raw Data -> Publishing Data

Raw Data -> UM/DM -> AM -> Anonymized Data -> Publishing Data


SecGraph: Evalua=on •  Datasets

–  Enron (36.7K users, 0.2M edges): an email network dataset

–  Facebook – New Orleans (63.7K users, 0.82M edges): a Facebook friendship dataset in the New Orleans area

•  Evalua=on senngs –  Follow the same/similar senngs in the original papers

–  See details in the paper and hDp://www.ece.gatech.edu/cap/secgraph/

•  Evalua=on scenarios –  Anonymiza=on vs U=lity

–  DA performance

–  Robustness of DA aDacks

–  Anonymiza=on vs DA


SecGraph: Anonymity vs U=lity

C1: existing anonymization techniques are better at preserving graph utilities

C2: all the anonymization techniques lose one or more application utility


SecGraph: DA Performance

0.00%

20.00%

40.00%

60.00%

0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95

De-‐anonymize Enron

NS DV JLSB

0.00%

20.00%

40.00%

60.00%

80.00%

100.00%

0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95

De-‐anonymize Facebook

NS DV JLSB

Phase transitional phenomena

C1: the phase transitional (percolation) phenomena is attack-dependent


SecGraph: Robustness of DA ADacks % of error seeds

0%

10%

20%

30%

40%

4%

8%

12%

16%

20%

24%

28%

32%

36%

40%

44%

48%

% OF SEED ERROR

De-‐anonymize Enron

NS DV ADA

0% 20% 40% 60% 80% 100%

4%

8%

12%

16%

20%

24%

28%

32%

36%

40%

44%

48%

% OF SEED ERROR

De-‐anonymize Facebook

NS DV ADA

C1: global structure-based DA attacks are more robust to seed errors


SecGraph: Anonymiza=on vs DA

C1: state-of-the-art anonymization techniques are vulnerable to one or several modern DA attacks

C2: no DA attack is general optimal. The DA performance depends on several factors, e.g., similarity between anonymized and auxiliary graphs, graph density, and DA heuristics


SecGraph: Release & Support •  Website

–  hDp://www.ece.gatech.edu/cap/secgraph/

–  Sopware

–  Datasets

–  Documents

–  Demo

–  Q&A

•  Modes –  GUI

–  Command line

•  Suppor=ng –  Windows

–  Linux

–  Mac


Conclusion

–  We analyzed exis=ng graph anonymiza=on techniques and de-‐anonymiza=on aDacks

–  We implemented and evaluated SecGraph: a uniform and open-‐source evalua=on system from graph data anonymiza=on and de-‐anonymiza=on

Acknowledgement We thank the anonymous reviewers very much for their valuable

comments!

We are grateful to the following researchers in developing SecGraph: Ada Fu, Michael Hay, Davide Proserpio, Qian Xiao, Shirin Nilizadeh,

Jing S. He, Wei Chen, and Stanford SNAP developers


Thank you! Shouling Ji

[email protected]

hDp://www.ece.gatech.edu/cap/secgraph/

hDp://users.ece.gatech.edu/sji/

secgraph:*a uniform and open-source evaluation system for ...secgraph:*a uniform and open-source...

Documents

secgraph:a uniform and open-source evaluation system for ...secgraph:a uniform and open-source...