SLIDE 1
Graph sharing
2
Towards Plausible Graph Anonymization Yang Zhang, Mathias Humbert, - - PowerPoint PPT Presentation
Towards Plausible Graph Anonymization Yang Zhang, Mathias Humbert, - - PowerPoint PPT Presentation
Towards Plausible Graph Anonymization Yang Zhang, Mathias Humbert, Bartlomiej Surma, Praveen Manoharan, Jilles Vreeken, Michael Backes Graph sharing 2 Graph anonymization 3 Graph anonymization id 3 id 7 id 1 id 6 id 4 id 2 id 8 id 5
SLIDE 2
SLIDE 3
Graph anonymization
3
SLIDE 4
Graph anonymization
4
id 6 id 2 id 1 id 7 id 8 id 5 id 4 id 3
SLIDE 5
Graph anonymization
5
id 6 id 2 id 1 id 7 id 8 id 5 id 4 id 3
SLIDE 6
Graph anonymization
6
id 6 id 2 id 1 id 7 id 8 id 5 id 4 id 3
SLIDE 7
Graph anonymization
7
id 6 id 2 id 1 id 7 id 8 id 5 id 4 id 3
SLIDE 8
Our work
▪ Find a fundamental flaw in graph anonymization designs
8
SLIDE 9
Our work
▪ Find a fundamental flaw in graph anonymization designs ▪ Exploit it to recover original graph
9
SLIDE 10
Our work
▪ Find a fundamental flaw in graph anonymization designs ▪ Exploit it to recover original graph ▪ Use our findings to enhance anonymization designs
10
SLIDE 11
Our work
▪ Find a fundamental flaw in graph anonymization designs ▪ Exploit it to recover original graph ▪ Use our findings to enhance anonymization designs ▪ Evaluate privacy and usability of enhanced techniques on 3 real life datasets: ▪ Enron, NO, Snap
11
SLIDE 12
Graph anonymization methods
▪ ’08 Liu et al. - k-anonymity (k-DA) ▪ ’08 Zhou et al. - k-anonymity (k-NA) ▪ ’10 Cheng et al. - k-anonymity (k-iso) ▪ ’11 Sala et al. - differential privacy ▪ ’12 Mittal et al. - random walk privacy ▪ ’14 Xiao et al. - differential privacy
12
SLIDE 13
k-DA algorithm
13
id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2
SLIDE 14
k-DA algorithm
id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2 # nodes
1 2 3 4 5
node degree
1 2 3 4
SLIDE 15
k-DA algorithm
15
2-DA id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2 # nodes
1 2 3 4 5
node degree
1 2 3 4
# nodes
1 2 3 4 5 6
node degree
1 2 3 4
SLIDE 16
k-DA algorithm
16
2-DA id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2 id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2 # nodes
1 2 3 4 5
node degree
1 2 3 4
# nodes
1 2 3 4 5 6
node degree
1 2 3 4
SLIDE 17
SalaDP algorithm
17
ɛ-DP
dK-2 series perturbed dK-2 series
id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2 id 6 id 4 id 1 id 8 id 7 id 5 id 3 id 2
SLIDE 18
Social network graph properties
18
id 6 id 2 id 1 id 7 id 8 id 5 id 4 id 3
SLIDE 19
id 2 id 1 id 3 id 5 id 4
Social network graph properties
19
id 6 id 7 id 8
SLIDE 20
id 4
Social network graph properties
20
id 6 id 2 id 1 id 7 id 8 id 5 id 3
SLIDE 21
Social network graph properties
21
id 6 id 2 id 1 id 7 id 8 id 5 id 4 id 3
SLIDE 22
Graph recovery attack - overview
22
SLIDE 23
Graph recovery attack - graph embedding
23
▪ Node embeddings with node2vec ’16 Grover and Leskovec ▪ Mapping users into continuous vector space ▪ User’s vector reflects structural properties
SLIDE 24
Graph recovery attack - graph embedding
24
▪ Plausibility is cosine similarity between embeddings
−0.2 0.0 0.2 0.4 0.6 0.8 1.0
Edge plausibility
1 2 3 4 5 6 7
Number of edges
×104 Original edges Fake edges
SLIDE 25
Graph recovery attack - graph embedding
25
▪ Plausibility is cosine similarity between embeddings
−0.2 0.0 0.2 0.4 0.6 0.8 1.0
Edge plausibility
1 2 3 4 5 6 7
Number of edges
×104 Original edges Fake edges
Enron NO SNAP 0.0 0.2 0.4 0.6 0.8 1.0
AUC
Cosine Euclidean Bray-Curtis Embeddedness Jaccard Adamic-Adar
SLIDE 26
Graph recovery attack - graph embedding
26
▪ Find a cutoff point and remove non-plausible edges
−0.2 0.0 0.2 0.4 0.6 0.8 1.0
Edge plausibility
1 2 3 4 5 6 7
Number of edges
×104 Original edges Fake edges F1 score
SLIDE 27
Enhancing anonymization
▪ get fake edges with highest plausibility? ▪ the distribution will look unnatural
27
SLIDE 28
Enhancing anonymization
▪ get fake edges with highest plausibility? ▪ the distribution will look unnatural ▪ draw fake edges from same plausibility distribution?
28
SLIDE 29
Enhancing anonymization
▪ get fake edges with highest plausibility? ▪ the distribution will look unnatural ▪ draw fake edges from same plausibility distribution?
29
k-DA (k=100) Enhanced k-DA (k=100)
SLIDE 30
Resilience to graph recovery attack
▪ F1 score for original anonymizations ▪ F1 score for enhanced anonymizations
30
k-DA drops by: 26~51% SalaDP drops by: 37~48%
SLIDE 31
Utility of Enhanced anonymization
31
0.6 0.7 0.8 0.9 1.0
Utility of GA
0.6 0.7 0.8 0.9 1.0
Utility of GF
Eigencentrality (Enron) Eigencentrality (NO) Eigencentrality (SNAP) Degree distribution (Enron) Degree distribution (NO) Degree distribution (SNAP) Triangle count (Enron) Triangle count (NO) Triangle count (SNAP)
SLIDE 32
Resilience to deanonymization attack
32
Enron NO SNAP 5 10 15 20 25 30
Anonymity gain (%)
k-DA (k =50) k-DA (k =75) k-DA (k =100) SalaDP (✏ =100) SalaDP (✏ =50) SalaDP (✏ =10)
SLIDE 33
Conclusion
33
We find flaws in current graph anonymizations
SLIDE 34
Conclusion
34
We find flaws in current graph anonymizations We recover the original, pre-anonymized graph
SLIDE 35
Conclusion
35
We find flaws in current graph anonymizations We enhance the anonymization techniques We recover the original, pre-anonymized graph
SLIDE 36
Conclusion
36
We find flaws in current graph anonymizations We enhance the anonymization techniques We evaluate privacy and utility
- f enhanced anonymization
We recover the original, pre-anonymized graph