Slides - Computer Science and Engineering

Report
Aegis: A Semantic Implementation of
Privacy as Contextual Integrity in
Social Ecosystems
Imrul Kayes, Adriana Iamnitchi
Social Privacy Risks
2
Why Does This Happen?
• Inappropriate sharing and transferring of
information
• (Permissive) Default privacy settings by OSN
provider
• Because they can
• Lack of universal framework that establishes
what is right and wrong
• Users do not change default settings
• 99% Twitter users
• >80% Facebook users
• When they do, they get it wrong
3
Evolution Towards Social Ecosystems
Applications
Social Inference API
Social Data Management
Personal Aggregators
Social Sensors
Social Signals
Iamnitchi et al. ”The Social Hourglass: an Infrastructure for Socially-aware
Applications and Services." IEEE Internet Computing (2012).
4
Privacy in Social Ecosystems
• Social Ecosystems amplify privacy concerns
– Aggregated data from different contexts of
activity
– A more complete (uncomfortable?) digital
recording of a person’s life
– Social applications from different contexts of
activity
• Default privacy settings become critical
5
Privacy as Contextual Integrity
• The right to appropriate flow of personal information
• Based on two life facts:
– transfer of personal information happens in a social
context
– people alter behavior to correspond with the norms of
the context
• Two norms:
– Norms of appropriateness
– Norms of distribution
Nissenbaum, Helen. "Privacy as contextual integrity." Washington Law
Review 79.1 (2004).
6
Our Solution
• Ontology-based social ecosystem data model to
capture user online data semantics
– Model social contexts
– Model user roles
• Generate default privacy from social data based
on Nissembaum’s contextual integrity framework
• Extensible, fine-grained default policy
customizable by users
• Prototype implementation and experimental
evaluation on three real-world large networks
7
e,
of
a
of
e
a
y
y
p-
al
d
s,
y
new social signals (see next section): the developers of social
sensors have to be aware of the ontology of the social contexts
Ontology-based
Social
Data
to
which the sensors report, in
order toEcosystems
maintain structural
data
representation.
Another way to extend the social ecosystem is
Model
by extending a social context itself when new relevant social
• Set ofbecome
entities, instances,
relations
and axioms
signals
available: functions,
for example,
Facebook
recently
• A vocabulary
for social
ecosystems
added
a service called
Gifts
which allows users to buy presents
• Provides formal and structured representation of user’s data
for their friends. Consequently, the social context model needs
and social spheres
to
be adaptive
to interoperability
accommodate additions of new contexts.
• Gives
semantic
Ontologies
designing
a scalable context model.
• High-levelhelp
logicininference
is possible
Context
Upper
Ontology
Friendship
Context
isa
Professional
Context
Gaming
Context
Blogging
Context
'Other'
Context
8
Context
specific
ontologies
es
phi
Tro
ges
Bad
s
ard
e ar
ns
ers'
Gamoup
Gr
es
Gam
ea
isT
m
ns
m
joi
Content
owned
Games
son
Per
Comment
O
ate
isI
n
f
Per
son
Experience
Industry
Job
Professional
Context
profe
s
Academia
Pro
fes
Gro siona
up l
Note
Event
Check
edIn
isF
rie
ndO
f
isL
oc
ate
d
Photo
In
Video
tI
sho
Place
Person
on
ndati
mme
o
c
e
hasR
isC
f
eO
gu
a
e
oll
ber
ce
rien
xpe
E
s
ha
siona
lMem
ha
sJ
ob
Person
s
attend
Group
tak
en
In
Aw
in g
Rac
zle
Puz
y
teg
Stra
cr
vo eate
lv e s
dI
n
Gaming
Context
Friendship
Context
Recommendation
Class
owl: Property
rdfs: subClassOf
Legend:
9
n
System Model
• Unrestricted set of disjoint social contexts
• A user belongs to only one social context at any time
• A user can have one or more roles in every social context s/he is
part of
• Each piece of data (resource) is assigned (created) to only one
context
• Shared data(resources) are replicated in each of the other users’
current contexts
• A request for a resource is made on behalf of the requester’s role
in the particular context in which the requester is when the
request is made
• A request specifies an action, which could be read, write, delete
or replicate to another user’s ownership.
10
Photo
Video
tIn
extracting and applying the default policies as well. This
component communicates with the Social Data Management
Layer which implements social contexts and roles.
Architecture
or is an
vice as
serving
Social
Data
Extractor
Policy
Evaluator
Default
Policy Repository
User
Defined
Policy
Extractor
Aegis: Privacy
Management
Layer
Contextual
Policy
Definer
Ontology
SEKB
User
Social Data
Management Layer
Ontology
Social Data Acquisition
&
Aggregation Layer
Policy Editor
ntology
exts. A
specific
Socially-aware Applications
A1
S11
S21
S22
A2
S32
A3
S33
S43
11
Policy Specification
• A policy is defined as a set of RDF statements
• Policies obey the two information norms of CI
Norms of appropriateness: Bob’s colleagues can read his professional
groups in the Professional context
Yes
Professional Groups?
Alice
Colleagues
ASK
where {
?req rdf:type p:requestor.
?req p:allowed p:read.
p:read p:performedOn Bob.
?req se:isColleagueOf Bob.
Bob se:professionalMember
?group.}
Bob
Charlie
12
Policy Specification
Norms of distribution: policy restricts the access to Bob’s photos
if they are shared
Shared contents
(e.g., Photo)
Alice
friends
ASK
where {
?req rdf:type p:requestor.
?req p:allowed p:read.
p:read p:performedOn Bob.
?req se:isFriendOf Bob.
Bob se:hasPhoto ?photo.
?photo se:status
se:notShared}
Bob
Charlie
13
p. }
personalized policy will be enforced (see policy evaluation
flow chart from Figure 4). As in the data model we have
hierarchy among classes (which eventually define resources)
and a group of classes belong to a context, we can infer
the context from requested resource. Note that a default
auto generated policy could be personalized by the user and
in this case, defines
the personalized
policyamong
will be evaluated.
For(user
Ontology
hierarchy
resources
example, from a requested resource recommendation, a context
data)
inference is possible from the following SPARQL query to
Context
is possible for each resource
knowledgeinference
base:
Context Inference
•
•
emove auxiliary
formation from
SEKB
Deny
access
act> , where
n instance of
PREFI X r df : <ht t p: / / www. w3. or g/ 1999/ 02/ 22r df - synt ax- ns#
PREFI X r df s: <ht t p: / / www. w3. or g/ 2000/ 01/
r df - schema#
PREFI X se: <ht t p: / / www. dsg. cse. usf . com/ se>
SELECT ?super Cl ass
Wher e {
se: r equest edResour ce r df s: subCl assOf
?super Cl ass . }
The policy engine will extract the policy of the inferred context
and execute it.
14
Al i ce p: al l owed p: r ead.
p: r ead p: per f or medOn Bob.
Al i ce se: i sCol l eagueOf Bob.
Bob se: pr of essi onal Member ?gr oup. }
Our policy rep
for each resourc
infer the contex
personalized pol
flow chart from
hierarchy among
and a group of
the context from
auto generated p
in this case, the
example, from a r
inference is poss
knowledge base:
Request Handling Flow Chart
Accept request
(requestor, resource, action)
Add request specific
auxiliary information
to SEKB
Default
Policy?
Yes
Infer context from
the requested
resource
Extract policy
No
Extract
policy
Execute
Query
Grant
access
Yes
Triple
returned?
remove auxiliary
information from
SEKB
No
Deny
access
Fig. 4: Request handling process.
15
PREFI X r df : <
r df - synt ax- n
PREFI X r df s:
r df - schema#
PREFI X se: <h
SELECT ?supe
Wher e {
se: r eque
?super Cl ass
Prototype Implementation
• Implemented the prototype in Java Platform
Standard Edition 6 (Java SE 6)
• Jena’s APIs for RDF data management
• Ontology: Jena’s API for handling OWL
ontologies
• leveraged TDB for persistent storage of
knowledge base
• SPARQL: Jena’s query engine
16
Experimental Evaluation
• Objective:
– Performance of the policy engine in
executing default policies for realistic
workloads
– Scalability of the policy engine in executing
default policies
– Overhead induced by default policies
17
Experimental Evaluation
• Three real networks
• Thirteen test cases (100~70,000 users): snowball
sampling from the networks
• Social ecosystems knowledge base including Person,
Relationships and Groups
• Two types of responses
- positive authorization access control response
- negative authorization access control response
18
Access time increases linearly with
the size of the SEKB
Positive authorization
Negative authorization
19
Number of requests answered per second
Positive and negative authorization
take about the same time
• TDB data structures are
threaded B+Trees
• long scans (negative
authorizations) proceeds
without needing to
traverse the branches of
the tree
20
Performance decreases with
increasing users
• Increased system
memory to realistic
capacity for an
in-production server
• Distributed solutions for
data management
21
Overhead induced by default policies is
Statistically Insignificant
Time taken(ms)
10
4
10
3
10
2
Slashdot
Slashdot with
BlogCatalog
BlogCatalog with
Facebook
Facebook with
without
default
without
default
without
default
policy
policy
policy
policy
policy
policy
101
100
10
0
50
0
10
00
25
00
50
0
0
75
00
10
00
0
20
0
00
30
00
0
40
00
0
Number of users in SEKB
22
50
00
0
60
00
0
70
0
00
Future Work
• Test the effects of default policies
- on applications that are too restrictive
- user satisfaction with user-based
surveys
• Formalize and analyze potential privacy
attacks
• Understand the system in different
platform settings
23
Summary
• Propose an ontology-based social ecosystem
data model to capture user social data
• Employ semantic web technologies to generate
default privacy polices based on Nissembaum’s
contextual integrity theory
• Provide an architecture and prototype
implementation of privacy model
• Experimental evaluation on three real-world
large networks to demonstrate the applicability
in practice
24
Thank You!
Aegis: A Semantic Implementation of Privacy
as Contextual Integrity in Social Ecosystems
Imrul kayes, Adriana Iamnitchi
http://www.cse.usf.edu/dsg/
[email protected]
25
Back Up Slides
26
Social Sensors
Consume social signals:
• Location/collocation
• Schedule (Google calendar)
• Mobile phone activity (calls, etc)
• Online social network interactions
• Email
• Shared content (Netflix, CiteULike)
• Personal relations (family)
…
27
Social Sensors
• Report on behalf of ego:
– Alter, the person ego is interacting with
– An activity tag: e.g., “outdoors”, “dining”
• Based on content, location, predefined labels,
semantic web (ontologies), etc.
– A weight: e.g., 0.15
• Run on ego’s mobile devices, desktop, or on
the web
• Process user interactions
– To reduce noise
– To distinguish between routine and meaningful
interactions
28
Aggregators
• Act as the user’s personal assistant
• Runs on trusted device (cell phone)
• Responsible for
o Managing access to social
signal apps
o Personalization
o Identity management
29
Related Work
• Squicciarini et al. “PriMa”
– auto generates access control policies for users
– Based on factors such as average privacy
preference of similar and related users, accessibility
of similar items in similar and related users,
closeness of owner and access or popularity of the
owner
– A large number of factors and their parametrized
tuning is required
– No performance evaluation
30
Related Work
• Shehab et al. “PolicyMgr”
– leverages user provided example policy
settings as training sets and build
classifiers that are the basis for autogenerated policies
– Practicality in terms of response time has
not yet been shown
31
Related Work
• Our privacy model differs from other
solutions
– We focused on generating default policies for
a social ecosystem that deals with users’
aggregated social data from different domains
– We considered a privacy framework proposed
by social theorists and translated it into an
architecture and proof-of-concept
implementation
32

similar documents