Caché: Caching Location-Enhanced Content to Improve User Privacy

Carnegie Mellon University

Shahriyar Amini, Janne Lindqvist, Jason HongJialiu Lin, Eran Toch, Norman Sadeh

June 30, 2011

Widespread Adoption of Location-Enabled Devices

2

• 2009: 150M GPS-equippedphones shipped

• 2014: 770M GPS-equippedphones expected to ship (~5x increase!)

• Future: Every mobile device will be location-enabled (GPS or WiFi)

[Berg Insight 2010]

Apps Reveal Private Information• App reveals:

– Time of Use– User Interest– Current Location

• Over time:– Mobility Behavior– Significant Locations– Socioeconomic Status?

3

Our Approach• Pre-Fetch content for large regions• Store content on mobile device• Respond to queries using stored content• Periodically update content

4

Pre-Fetching Insight• Some location-based content

are still useful even when old (time to live)

5

Update Rate

Data Type

Real-Time(STTL)

Traffic flow, parking spotse.g. Loopt, PeopleFinder, Reno, Bustle

Daily weather forecasts, social events, couponse.g. Dede

Weekly movie/theatre schedules, advertisements, crime ratese.g. Yelp!, GeoNotes, PlaceIts, PlaceMail

Monthly restaurant guides, bus schedules, geocachese.g. Wikipedia (geo-tagged pages)

Yearly maps, points of interests, tour guides, store locatorse.g. Google Maps, Starbucks, Wal-Mart

LongTime

ToLive

Feasibility of Pre-Fetching• Content doesn’t change too often

– Average daily amount of changeover a 5 month period

• Requires <20 MB for Pittsburgh, ~100 MB for NYC

6

Pre-fetching Content• Geo-coordinates are continuous• Content cannot be pre-fetched for every point• Use a grid to discretize space

7

8

Caché Architecture

Application Design• Developer defines:

– Size of cells– Content update rate– Query string

9

http://api.yelp.com/v2/search?term=food&ll=#SLL_LAT#,#SLL_LON#

Application Installation• Regions of interest:

– 15213– Pittsburgh, PA

• Pre-fetch radius– 1 km

10

Content Download• Pre-fetch only when:

– Plugged in– Connected to WiFi

• Pre-fetch every cell

• Update content at defined update rate

11

Content Retrieval• Assume fresh content

• Retrieve content from a single cell

• Content miss results in a live request to LBS

12

High Content Hit Rate!• Evaluation Based on Mobility Datasets

– Locaccino: location-sharing– Place naming: label location

13

Radius (miles) Locaccino Place Naming

5 86% 79%

10 87% 84%

15 87% 86%

Caché Android Service• Android background service for apps

– Apps modified to make requests to service

– Service only pre-fetches when device is plugged in and connected to WiFi

14

Limitations• Doesn’t work for

– Rapidly changing content (STTL)– Apps with client/server interaction (Facebook)– Apps with server computation (Navigation)

• Burden falls on the developer– Developer responsible for defining cell size and

content download parameter• New regions have to be specified

before use

15

Our thoughts• Merit: interesting method to protect privacy• Cannot provide absolute privacy• Hard to switch between privacy-protected

mode and normal mode

16

ConclusionThe most private request is the one you don’t make.

17

• Feasibility analysis: STTL and LTTL, storage• Implement the privacy protected system• Two evaluations: human mobility dataset

and open source application

Locaccino• December 2008 to October 2009• Top 20 users• 5 female• Grads, Undergrads, Faculty, Research Dev

18

Place Naming• Ages 19 to 46• Spring, Summer, and Fall 2009• 45% female in Spring (N = 33)• 40% female in Summer (N = 26)• 30% female in Fall (N = 10)

19

Study of Mobility Traces• How often will queries be cached?

– Locaccino: Top 20 people, 460k traces– Place naming: 26 people, 118k traces

– Pre-Fetch newly visited regions over night

20

Radius (miles) Locaccino Place Naming

5 86% 79%

10 87% 84%

15 87% 86%

Radius (miles) Locaccino Place Naming

5 96% 79%

10 97% 85%

15 97% 87%