Privacy Preserving Data Mining

1
Healthcare privacy and security Genomic data
privacy
Li Xiong CS573 Data Privacy and Security
2
Genomic data privacy

• Genomic data are increasingly collected, stored,
  and shared in research and clinical environments
• Genomic data are person-specific (there exists no
  public registrar that maps genomes to names of
  individuals)
• Genomic data is not specified as an identifying
  patient attribute under HIPAA privacy rule and
  may be released for public research purposes
• How can person-specific DNA be shared, such
  that it cannot be associated to its explicit
  identity?

3
Data sharing scenario

• John Smith admitted to a local hospital which
  stores clinical and DNA information
• John visits other hospitals
• The hospital forward certain DNA data onto a
  research group, with institution and pseudonyms
  of the patients
• The hospital sends identified discharge record
  onto a state-controlled database

4
Data at a specific location

• Identified table of patient demographics
• De-identified DNA sequences
• Can we uniquely link identified data to DNA data?

5
Data at multiple locations

• Each site has an identified table and
  de-identified DNA sequences
• Can we uniquely link identified data to DNA data?

6
Trails

• The set of locations each patient visited is
  called a trail
• The trails can be tracked and matched to link DNA
  data to identified data

7
REIDIT-Complete

• Re-identification of data in trails (REIDIT) for
  complete publishing
• If there is a unique trail match, then a
  re-identification occurred

8
Results
9
REIDIT-C reidentification

• Re-identifiability related to average people
  per location

10
Reserved publishing

• Data releasers can reserve certain information
• N is reserved to P vs. P is reserved to N

11
REIDIT - Incomplete

• REIDIT for reserved publishing
• For each trail in the track with incomplete
  trails, if there is only one supertrail, then a
  re-identification occurred
• Remove the re-identified supertrail
• Important because a trail can be a supertrail to
  many trails
• Repeat the process

12
REIDIT-Incomplete
0.0, 0.1, 0.5, 0.9 probability of reserving
information hospital rank based on of patients
13
Can masking location help?
Not necessarily!
14
Comments and open issues

• Can k-anonymity solve the problem?
• Pseudonyms subject to dictionary attacks, how to
  allow linkage of the data without pseudonyms
• Genomic protection methods incorporating utility
  of the genomic data

15
(No Transcript)
16
De-identification
e.g. Utah Resource for Genetic and Epidemiologic
Research (RGE)