Privacy Decision Making in IoT Scenarios Pardis Emami-Naeini , Martin - - PowerPoint PPT Presentation

The Influence of Friends and Experts on Privacy Decision Making in IoT Scenarios

Pardis Emami-Naeini, Martin Degeling*, Lujo Bauer, Lorrie Cranor, Mohammad Reza Haghighat†, Richard Chow†, Heather Patterson†

* †

Alice in Wonderland…

2

Data: temperature Retention: 1 day Purpose: adjust room’s temperature Fewer than 15% of privacy experts allowed More than 65% of your friends allowed Data: fingerprint Retention: forever Purpose: authentication Data: video Retention: 1 month Purpose: security

Alice’s Phone

social influence

What is the impact of social influence on people making privacy-related decisions about allowing data collection by IoT devices?

Rese esearch qu questio ion

3

Soc

• cia

ial l in influence

Intentional or unintentional changes to individuals’ opinions or behaviors caused by others.

4

No! Yes!

We e stu tudied in indir irect, informatio ional soc socia ial in influence

normative informational direct indirect

5

Nor

• rmativ

ive or

• r in

informatio ional soc socia ial in influ luence

6

informational normative

We e stu tudied in informatio ional soc socia ial l influence

7

informational

Dir irect or

• r indir

irect soc social l in influ luence

8

direct indirect

We e stu tudied in indir irect soc social l in influ luence

9

indirect

Desi esigned a a vig ignette stu tudy

Short hypothetical stories…

10

Once upon a time …

Sce cenario ios wher ere be benefit its ou

• utweig

igh risk risks

11

Sce cenario ios wher ere risk risks ou

• utweigh be

benefit its

12

Sce cenario ios wit ith a a ba bala lance

13

Used sed a a pr pre-study to

• pi

pick sce scenario ios

• 500 Mechanical Turk participants
• From the United States
• Presented with 28 hypothetical IoT data-collection scenarios
• Asked whether participants would allow data collection
• Compensated for $2.50

14

Sele elected 9 pr pre-study sce scenario ios

15

3 allow more than 80% allowed fewer than 20% allowed 45% to 55% allowed 3 deny 3 balanced

More than 85% of privacy experts allowed

Con

• nsis

istency

16

Consistent

Allow!

Data: temperature Retention: 1 day Purpose: adjust room’s temperature

Alice’s Phone Pilot Participants

Con

• nsis

istency

17

Inconsistent Data: temperature Retention: 1 day Purpose: adjust room’s temperature

Alice’s Phone

Fewer than 15% of privacy experts allowed

Allow!

Pilot Participants

Two

• con
• nsensus le

level l for

• r soc

socia ial cue cues

• “More than 85% of [influencer] allowed the data collection.”
• “Fewer than 15% of [influencer] allowed the data collection.”
• “More than 65% of [influencer] allowed the data collection.”
• “Fewer than 35% of [influencer] allowed the data collection.”

18

strong weak

5 stu tudy con

• ndit

itions

• Out of 9 scenarios in each experimental condition
• 5 strong social cues
• 4 weak social cues

19

inconsistent experts control condition consistent experts consistent friends inconsistent friends consistent inconsistent inconsistent consistent

Ex Exam ample le of

• f a

a bal balanced sce scenario io

You are at the library. This message is displayed

• n your smartphone:
• Your smartwatch is keeping track of your

specific position.

• Your position is used by the smartwatch to

determine possible escape routes in the case

• f an emergency.
• This data will never be deleted.
• [Experimental conditions] Fewer than 35% of

your friends allowed this data collection.

20

1000 Mec echanic ical Tur urk par partic icipants

• From the United States
• 200 participants per condition
• Avg. age: 35
• ~15 minutes to complete
• Compensated for $2.50

21

Que uestio ions pe per sce scenario io

• If you had the choice, would you allow or deny this data collection?

22

inferred impact of social influence

Que uestio ions aft fter 9 all all sce scenarios

• When considering the 9 scenarios above, how much were you

influenced by the decisions that [influencer] made in these scenarios?

23

self-reported impact of social influence

Used sed reg egressio ion to

• ana

analy lyze

• Applied GLMM + random intercept
• Model selection by backward elimination

24

Soc

• cia

ial l in influence e mak akes a a dif difference!

• People are influenced by privacy experts and their friends differently
• Example: 11% more allowed in the “allow” scenarios when influenced by

consistent experts, compared to control condition with no influence

• Social influence speeds up decision making

25

Sure!

Soc

• cia

ial l in influence e spe speeds up up de decisio ion mak akin ing

• In general among all conditions:
• allow < deny < balanced
• Impact of social influence:
• With social influence (3.69 s) < without social influence (4.24 s)
• Biggest impact on balanced scenarios:
• With social influence (3.61 s) < without social influence (4.55 s)

26

Fac actors im impactin ing the the ext xten ent t of

• f soc

socia ial l in influence

• Task difficulty
• Most influence in balanced scenarios
• Consistency
• Consistent social cues have more influence
• Strength of social cues
• Strength of cues directly relates to the influence
• Type of influencer
• Experts allow
• Friends deny

27

Reported to

• be sign

ignificantly more in influenced by con

• nsistent frie

friends th than by in inconsistent frie friends

28

Influenced to oppose the cue Influenced to follow the cue

not influenced

inconsistent social cue

200 150 100 50 number of participants

Reported to

• be sign

ignificantly more in influenced by con

• nsistent frie

friends th than by in inconsistent frie friends

29

Influenced to oppose the cue Influenced to follow the cue

not influenced not influenced

200 150 100 50 number of participants

consistent social cue inconsistent social cue

Reported to

• be sign

ignificantly more in influenced by con

• nsistent exp

xperts th than by in inconsistent experts

30

Influenced to oppose the cue Influenced to follow the cue

not influenced

number of participants

not influenced not influenced

consistent social cue inconsistent social cue

200 150 100 50 number of participants

Peo eople le rep eported to

• pr

prefer in influence fr from

• m exp

xper erts

• Reported to be significantly more influenced when being asked about

privacy experts in control condition

31

Influenced to oppose the cue Influenced to follow the cue cue comes from friends

200 150 100 50 number of participants

not influenced

Peo eople le rep eported to

• pr

prefer in influence fr from

• m exp

xper erts

• Reported to be significantly more influenced when being asked about

privacy experts in control condition

• Most mentioned quality to be influenced by: having background in

technology

32

Influenced to oppose the cue Influenced to follow the cue

not influenced

200 150 100 50 number of participants

not influenced

cue comes from friends cue comes from experts

not influenced

Soc

• cia

ial l in influence e in in act actio ion

• Social influence is a promising approach for privacy assistants
• Important to choose influencers carefully and evaluate them over time

Pardis Emami-Naeini, Martin Degeling, Lujo Bauer, Lorrie Cranor, Mohammad Reza Haghighat, Richard Chow, Heather Patterson More info: www.privacyassistant.org

33