Towards Exploring User Perception of a Privacy Sensitive Information

Detection Tool

Vanessa Bracamonte

, Welderufael B. Tesfay

and Shinsaku Kiyomoto

KDDI Research, Inc., Saitama, Japan

Goethe University Frankfurt, Frankfurt am Main, Germany

Keywords:

Privacy Tools, User Perception, Privacy Sensitive Information.

Abstract:

Users reveal privacy sensitive information when they post on social media, which can have negative conse-

quences. To help these users make informed decisions, different tools have been developed that detect privacy

sensitive information (PSI) and provide alerts. However, how users would perceive this type of tool has not yet

been evaluated. In this position paper, we take the ﬁrst steps to address this gap, by exploring user intention,

perceived usefulness and attitude towards a PSI detection tool. We designed an experiment and showed partic-

ipants examples of the PSI detection tool alerts, and quantitatively and qualitatively evaluated their response.

The results showed that participants perceived the PSI detection tool as useful, had positive interest, and a low

level of concern about it, although they had a neutral level of intention of using the tool. The participants’

open-ended responses revealed that they considered the PSI detection tool useful, but mostly for other people

and not for themselves. In addition, they were concerned about the privacy risks of using the tool and about its

effectiveness. The ﬁndings reveal the challenges that PSI detection tools have to overcome to gain acceptance

among users that would beneﬁt from this type of privacy protection.

1 INTRODUCTION

Over the last couple of decades, information technol-

ogy and digital services have become an integral part

of the online society. Always connected devices such

as the smartphone have fostered the transfer of many

ofﬂine services and activities to online. As a result,

internet users often release a huge amount of data

while using these services. As such, the user concerns

over the consequences of privacy and liberty have also

grown. To address these concerns, research efforts

have focused on devising privacy enhancing technolo-

gies, enacting data protection laws and studying user

information disclosure behaviours.

Regulations such as the EU General Data Pro-

tection Regulation (GDPR) aim to protect user’s pri-

vacy by enforcing different requirements. Privacy-by-

design and privacy default are among the key princi-

ples enshrined in the GDPR. In particular, Article 9

of the GDPR highlights that certain information types

need special care. Previous studies have demonstrated

that when users (un)intentionally divulge privacy sen-

sitive information (PSI), they often regret having done

so (Sleeper et al., 2013; Wang et al., 2011). Further-

more, Acquisti and Fong (2020) have shown that the

disclosure PSI such as information related to religious

afﬁliations and sexual orientation, can be utilized by

different parties to discriminate users, e.g., in job ap-

plications screening processes.

To help users make informed decisions with re-

spect to their PSI disclosures, Privacy Enhancing

Tools (PETs) such as Privacy Detective (Caliskan Is-

lam et al., 2014a) and PrivacyBot (Tesfay et al.,

2019) have been developed, which detect PSI in user-

generated unstructured texts. These tools have a

promising potential for empowering users; however,

users’ perception of these tools has not yet been eval-

uated. In this paper, our main objective is to explore

how users perceive a PSI detection tool, whether PSI

type and explanation information inﬂuences their per-

ception, and what are their most frequent comments

and concerns, in order to improve the design of these

tools.

2 RELATED WORK

This section gives an overview of the existing work

both in PSI detection in social networking sites, and

628

Bracamonte, V., Tesfay, W. and Kiyomoto, S.

Towards Exploring User Perception of a Privacy Sensitive Information Detection Tool.

DOI: 10.5220/0010319706280634

In Proceedings of the 7th International Conference on Information Systems Security and Privacy (ICISSP 2021), pages 628-634

ISBN: 978-989-758-491-6

user evaluations of Privacy Enhancing Technologies

(PETs) based on Machine Learning (ML) techniques.

2.1 PSI Detection

Detecting PSI, especially in social networking sites,

is a challenging yet growing area of research (Tes-

fay et al., 2016). Wang et al. (2011) presented two

models for the prediction of personally identiﬁable in-

formation (PII) in emails. Similarly, Bier and Prior

(2014) developed a process that focuses on the auto-

mated recognition of PII. The aim of their study was

to enable companies to uncover PII in incoming email

communication. Sokolova et al. (2009) focused on

health information. Their work presented a mecha-

nism to detect privacy sensitive health information.

Castillo and Chen (2016) proposed a transfer

learning approach to detect PSI in tweets. Mao et al.

(2011) demonstrated the application of classiﬁcation

techniques to identify three types of privacy leaks

in tweets; namely, revealing dates of vacation plans,

tweeting under the inﬂuence of alcohol, and revealing

medical conditions. Tesfay et al. (2019) presented the

PrivacyBot PSI detection tool that identiﬁes 14 differ-

ent information types that are deﬁned in Art. 9 of the

EU GDPR and Caliskan Islam et al. (2014b)’s work.

2.2 User Evaluation of ML-based

Nudging PETs

When users navigate through different conﬁgurations

in online services, they encounter a large number

of decision points. These decisions often have a

big impact on their privacy. Nudging tools based

on machine learning algorithms for PSI detection in

unstructured texts could help users with these de-

cisions, but user evaluation remains unattended re-

search arena, and therefore we do not yet understand

how these tools may be perceived.

In general, however, nudging research has shown

promising results in inﬂuencing users towards more

informed privacy-related decisions (Acquisti et al.,

2017). Vishwamitra et al. (2017) studied the effect

of privacy-enhancing obfuscation such as “blurring”

and “blocking” on user perceptions measured by im-

age satisfaction, information sufﬁciency, enjoyment,

and social presence. Bracamonte et al. (2019) found

that users have a positive interest in PETs such as pri-

vacy policy summarization tools (Tesfay et al., 2018),

and that adding explanation information to the re-

sults of these tools, such as highlighting privacy pol-

icy segments, can increase perceptions of trustworthi-

ness and usefulness, as well as intention of use (Bra-

camonte et al., 2020).

These studies demonstrate that users, in general,

have a positive perception of PETs. However, to the

best of our knowledge, there is a limited research ef-

fort in applying similar user evaluation studies for PSI

detection tools.

3 METHOD

3.1 Experiment Design

We designed the experiment to explore user percep-

tion of the PSI detection tool, as follows. We deﬁned

two factors, Explanation and PSI type, with two lev-

els each. For Explanation, the levels were No Expla-

nation (Control) and Explanation. For PSI type, we

chose two types from the list deﬁned by Tesfay et al.

(2019): Health and Family.

The combination of factors resulted in four exper-

imental conditions (between-subjects design), and we

prepared two examples of the PSI detection tool alert

for each of the conditions. The content of the social

media posts for the examples was related to Health

or Family, according to the deﬁnition by Tesfay et al.

(2019), and were taken from Twitter.

In the alert examples, we used highlighting of im-

portant words to simulate the explanation of the re-

sult of the PSI detection tool. Although existing PSI

detection tools do not currently provide explanations

for their results, the technique of keyword highlight-

ing for explanation visualization is commonly used

for text data. For the conditions with explanation, we

highlighted words in the text which were important

to its classiﬁcation as privacy sensitive within its type

(Health or Family). The conditions without explana-

tion (Control) used the same texts, but without high-

lights. Figure 1 shows how the alert examples were

presented to the participants in the survey.

3.2 Questionnaire

After viewing the privacy alert examples, we asked

the participants questions regarding intention of use,

perceived usefulness and attitude (interest and con-

cern about the tool) regarding the PSI detection tool.

Table 1 shows the detail of the questions. We included

an open-ended response question for the participants

to indicate their opinion about the privacy alert tool

(”Please write your opinions, comments or concerns

about the privacy alert tool described.”).

The questionnaire also included questions about

the participants’ social media use and social media

privacy concerns, to help characterize the sample.

We asked questions on frequency of social media

Towards Exploring User Perception of a Privacy Sensitive Information Detection Tool

629

Figure 1: Example of the privacy sensitive information de-

tection tool alert included in the questionnaire, correspond-

ing to the Health PSI type. (a) Base design (Control). (b)

Alert with highlighted words (Explanation).

Table 1: Questionnaire items for the main variables of inter-

est in this study. The responses are on a 7-point scale from

Strongly disagree to Strongly agree.

Variable Question

Intention I would use this tool to receive this

type of privacy alert.

Usefulness This tool would be useful to decide

whether to share privacy sensitive

information.

Interest I am interested in trying out this tool.

Concern I have concerns about trying out this

tool.

use (”I use social media (Twitter, Facebook, etc.)”)

and personal information posting on social media (”I

post personal information on social media (Twitter,

Facebook, etc.)”) with a 7-point response scale rang-

ing from Never to Very frequently. We included

a question on privacy concerns when using social

media (”I worry about the consequences of posting

personal information on social media”) with a 7-

point response scale ranging from Strongly disagree

to Strongly agree, and a question on whether the par-

ticipants used any privacy tools (”Do you currently

use any tools for protecting your privacy on social

media?” (Yes/No)). For this last question, we asked

participants who had responded Yes to indicate which

privacy tools they used.

Finally, the questionnaire included an attention

check question, and age and gender questions.

3.3 Participants

We conducted the study on the Amazon Mechanical

Turk (AMT) platform, on October 14-15, 2020. AMT

workers with the following characteristics were re-

cruited: had worked a minimum of 1000 tasks in the

platform (HITs), had an 99% task approval rate, and

were from the USA, Canada, Australia or the UK. The

participants were compensated with $1.25 for com-

pleting the task. While running the study, we iden-

tiﬁed responses with answers which were completely

unrelated to the questions and which failed the atten-

tion check question. These responses were not ap-

proved and were not included in the data.

4 RESULTS

4.1 Sample Characteristics

The survey initially obtained 160 responses. We ana-

lyzed the response to the attention check question and

identiﬁed 4 cases with incorrect answers. These cases

were eliminated from the analysis, which resulted in

a valid sample of 156 participants.

The gender distribution of participants was as fol-

lows (number of participants in parentheses): female

38.7% (60), male 60.6% (94), 1 other and 1 NA. The

age distribution was: 20-29 y/o 21% (33), 30-39 y/o

52% (81), 40-49 y/o 17% (27), 50-59 y/o 7% (11),

60+ y/o 2% (3), with 1 NA.

4.2 Social Media Use and Social Media

Privacy Concerns

As Figure 2 shows, the majority of participants fre-

quently used social media, with the highest proportion

corresponding to a Very frequently response.

Figure 2: Distribution of responses to the questions about

the participants’ frequency of social media use and personal

information posting, social media privacy concerns and use

of privacy tools. Numbers indicate percentage.

On the other hand, the majority of participants

indicated that they seldom posted personal informa-

tion on social media. Participants indicated concern

about the consequences of posting personal informa-

tion on social media: the median of responses was 6,

ICISSP 2021 - 7th International Conference on Information Systems Security and Privacy

630

signiﬁcantly greater than the neutral point (Wilcoxon

signed-rank test, p<0.001). The vast majority of par-

ticipants did not report the use of any privacy pro-

tection tools. Of the participants that indicated the

use of a privacy protection tool, 7 mentioned the use

of privacy settings in the social media platform itself

(e.g. ”only the privacy settings in the actual apps. like

friends can see posts only..”); 4 mentioned the use of

security settings (e.g ”password”); and 5 mentioned

the use of security/privacy tools (e.g ”VPN”, ”Norton

360”).

4.3 Perception of the Privacy Alert Tool

4.3.1 Quantitative Analysis

We used the non-parametric method Aligned Ranks

Transform ANOVA (Wobbrock et al., 2011) to test

whether the factors of Explanation and PSI type had

an effect on the variables of interest: intention of use,

perceived usefulness, interest and concern regarding

the PSI detection tool alert. The results of the analy-

sis indicated that there were no signiﬁcant effects of

either of the factors for any of the variable.

We proceeded to analyze the whole sample. Fig-

ure 3 shows the distribution of responses.

Figure 3: Distribution of responses to the questions about

the participants’ intention of use, perceived usefulness, in-

terest and concern regarding the PSI detection tool. Num-

bers indicate percentage.

Figure 4 shows the box plots with the median for

each of the variables of interest; the dashed line indi-

cates the neutral point.

We conducted the non-parametric one-sample

Wilcoxon signed-rank test, to evaluate whether the

scores for intention of use, perceived usefulness and

interest were signiﬁcantly greater than the neutral

score of 4. In the case of concern about the tool, we

tested whether the scores were signiﬁcantly less than

the neutral score.

The results indicate that the participants had a pos-

itive perception and attitude towards the tool: use-

fulness (p<0.001) and interest in the tool (p=0.001)

Figure 4: Box plots for intention of use, perceived useful-

ness, interest and concern regarding the tool. The dashed

line indicates the middle point of the response scale (neu-

tral).

were signiﬁcantly positive. In addition, participants

had a low level of concern about trying out the tool

(p<0.001). On the other hand, participants inten-

tion of use of the PSI detection tool was not positive

(p=0.11), with a median of 4 (neutral middle of the

scale) as can be observed in Figure 4.

4.3.2 Open-ended Comments

We qualitatively analyzed the content of the responses

to the open-ended question on the participants’ opin-

ions and concerns regarding the PSI detection tool.

We ﬁrst reviewed each response and classiﬁed them

by whether they contained all positive, all negative

or both positive and negative comments. During the

review we also identiﬁed responses that did not corre-

spond to the question (which were classiﬁed as Other)

and blank responses. Table 2 shows the results by ex-

periment condition.

The results show that in all conditions, the propor-

tion of participants that gave an all positive comment

was similar than those who gave an all negative com-

ment. The exception is the case of the Control-Family

condition, where the number of all positive comments

was more than twice the number of all negative com-

ments. The number of participants that mentioned

both positive and negative aspects was similar for all

conditions.

On the positive side, participants mentioned use-

fulness most frequently, followed by a general per-

ception of the tool being good or a good idea and in-

teresting (Table 3). With regards to usefulness, it is

interesting to note that participants frequently men-

tioned it as useful for others, not themselves.

On the negative side (Table 4), the aspects men-

tioned were more varied. Concern about the pri-

vacy risk of the PSI detection tool itself was most

frequently mentioned: participants were concerned

Towards Exploring User Perception of a Privacy Sensitive Information Detection Tool

631

Table 2: Classiﬁcation of open-ended comments from par-

ticipants, by condition. The table includes the category

Other (unrelated comments) and Blank (no answer).

Condition n %

Control-Health Positive 15 38%

Both 8 20%

Negative 13 33%

Other 3 8%

Blank 1 3%

Explanation-Health Positive 15 38%

Both 6 15%

Negative 16 41%

Other 1 3%

Blank 1 3%

Control-Family Positive 20 53%

Both 8 21%

Negative 8 21%

Other 2 5%

Explanation-Family Positive 12 31%

Both 10 26%

Negative 13 33%

Other 3 8%

Blank 1 3%

Table 3: Positive aspects of the PSI detection tool men-

tioned by participants in the open-ended comments.

Type n Example

useful 56 ”I think that this tool would

be useful to get people think-

ing about the type of things

they post online.”

good 22 ”I believe this is a good tool

to help those who are not

very aware of what sensi-

tive information they may be

posting online ”

interesting 11 ”I didn’t know it was a thing

so it’s interesting”

performance 2 ”Based on the examples, I’d

say this privacy alert tool is

doing its job.”

about how the PSI detection tool would handle their

private data and what it might do with it.

Participants mentioned the PSI detection tool not

being useful as the second most frequent negative as-

pect; speciﬁcally, participants considered that the PSI

detection tool was not useful for themselves, because

they already took the necessary precautions to avoid

posting personal information on social media. This

type of comment was frequently accompanied by pos-

itive comments regarding the potential usefulness of

the PSI detection tool for other people. The categories

of others that participants consider might beneﬁt from

Table 4: Negative aspects of the PSI detection tool men-

tioned by participants in the open-ended comments.

Type n Example

privacy 31 ” There is still worries about

that, if it analyses my data,

does it store it? If so the tool

itself is a privacy risk.”

not useful 21 ”However I myself would

likely have no use for it as I

don’t post everything about

my life like some others. ”

too sensitive 14 ”If the tool were conﬁg-

urable to be less sensitive

when it makes sense, it

could be handy without hav-

ing too many warnings”

performance 12 ”Concerns that I have on the

tool include how effective

the tool is in detecting pri-

vacy information and how

and why the tool detects cer-

tain words as sensitive.”

inconvenient 9 ”Also, I would ﬁnd it annoy-

ing after awhile I think.”

censorship 3 ”At ﬁrst glance, it just seems

like another form of censor-

ship. A tool attempting to

get us to censor ourselves.”

over-reliance 2 ”In fact, I’m kind of afraid

that having a reminder like

this might make me less

careful.”

monitoring 2 ” I feel like it’s just monitor-

ing me, which I don’t like.”

cost 1 ”(. . . ) but I would want to

know what I would have to

provide/download/etc to use

it.”

the PSI detection tool includes: children, young peo-

ple, older people, and in general people who are ”less

aware” of privacy risks on social media.

Next, participants were concerned about the PSI

detection tool being ”too sensitive”, and about the

general performance of effectiveness of the PSI de-

tection tool. In particular, the information related

to family and health was considered not sensitive by

some participants, and some of them made sugges-

tions about how the PSI detection tool could deal with

this issue. For example, they suggested having sensi-

tivity settings, whitelists or considering the context of

the post.

ICISSP 2021 - 7th International Conference on Information Systems Security and Privacy

632

5 DISCUSSION

The quantitative results shows a positive perception of

the usefulness of the PSI detection tool, and the qual-

itative analysis, where usefulness was the most fre-

quent positive aspect mentioned, supports this result.

Comments about usefulness were often accompanied

by phrases that provide evidence of the nudging ef-

fect of the PSI detection tool: ”useful to get people

thinking”, ”It would give them a second to pause and

think about if it is a good thing to post or not.”. On

the other hand, usefulness was also mentioned in neg-

ative terms and the examination of the content of the

responses reveal a dual nature in the participants’ per-

ception of usefulness of the PSI detection tool. In

summary, participants considered the tool useful, in

particular for others but not so much for themselves.

This contradiction can also explain the ﬁnding that the

level of intention of use was neutral and not signiﬁ-

cantly positive.

We cannot judge the accuracy of the participant’s

self-perceived privacy risk awareness, but this type of

response suggests that it would be beneﬁcial for PSI

detection tools to provide information on historical

data. This could be done through calculating privacy

risk scores (Aghasian et al., 2017), or by detecting ex-

amples from the user’s past social media posts where

(or if) PSI had been disclosed.

Participant’s comments show that the foremost

concern is the privacy risk of the PSI detection tool

itself. This type of privacy tool relies on an important

assumption that the users’ will grant access to their

private information, and users rightly wish to have

assurances that their data will not be misused. The

trustworthiness of the PSI detection tool in terms of

security and privacy protection should be established,

through technical means or reputable providers for

example, in order to allow users to safely grant it ac-

cess to current or historical data.

Participants’ concern regarding the tool being ”too

sensitive” also poses an interesting challenge for the

design of PSI detection. Current tools have a predeter-

mined classiﬁcation of what exactly constitutes ”sen-

sitive information”, but this classiﬁcation may not be

compatible with the users’ own concept of PSI. Re-

gardless of whether the user is correct, this incom-

patibility could result in false alarms (as perceived by

the user) and annoyance. PSI detection tools should

consider how to offer some ﬂexibility without com-

promising their goal.

Finally, we did not ﬁnd a signiﬁcant effect (pos-

itive or negative) of explanations or PSI type in the

perception of the tool. In the case of explanation, one

possibility may be that the explanations are superﬂu-

ous with such short texts. Future research will include

validation of these results.

5.1 Limitations

This study has the following limitations. First, the de-

sign of the study is not comprehensive, in the sense

that we did not include all possible PSI types as con-

ditions in the experiment. The main objective of the

study was to explore user perception and to identify

avenues of future research. Next, the data used to

construct the example alerts that we showed to par-

ticipants were taken from Twitter, a social media plat-

form that is characterized by short text length per post.

The responses from participants may be different for

alerts corresponding to longer or more complex so-

cial media posts, where the privacy sensitive informa-

tion contained in the text may be harder to judge at a

glance. Finally, we asked questions regarding opin-

ions and experience related to privacy sensitive and

personal information, but their exact deﬁnition may

be different for each participant. Future work will aim

to clarify and explore users’ understanding of these

concepts.

6 CONCLUSIONS

In this paper, we conducted a study to explore user

perception about a privacy sensitive information (PSI)

detection tool for social media. We showed partici-

pants examples of the alert provided by the PSI detec-

tion tool, with and without explanation of the result,

which warned them that they might be posting PSI

about health or family. We quantitatively and qualita-

tively evaluated the participants’ response.

The results showed that participants perceived the

PSI detection tool as useful, had positive interest, and

a low level of concern about it. On the other hand, par-

ticipants were neutral regarding whether they would

use the PSI detection tool, and we did not ﬁnd sig-

niﬁcant effects of explanations or between health and

family type of PSI for any of the variables. The analy-

sis of the open-ended comments indicated that partic-

ipants thought that the PSI detection tool was useful

(although mostly for others and not themselves), but

were concerned about the privacy risks posed by the

use of the tool itself. The participants’ comments also

included suggestions on possible features of the PSI

detection tool, which suggests that there was a level

of engagement with the idea. Future work is planned

to quantitatively validate the inﬂuence of the factors

identiﬁed in this study, and to apply the ﬁndings to

the design of a PSI detection tool.

Towards Exploring User Perception of a Privacy Sensitive Information Detection Tool

633

REFERENCES

Acquisti, A., Adjerid, I., Balebako, R., Brandimarte, L.,

Cranor, L. F., Komanduri, S., Leon, P. G., Sadeh, N.,

Schaub, F., Sleeper, M., et al. (2017). Nudges for pri-

vacy and security: Understanding and assisting users’

choices online. ACM Computing Surveys (CSUR),

50(3):1–41.

Acquisti, A. and Fong, C. (2020). An experiment in hiring

discrimination via online social networks. Manage-

ment Science, 66(3):1005–1024.

Aghasian, E., Garg, S., Gao, L., Yu, S., and Montgomery, J.

(2017). Scoring users’ privacy disclosure across mul-

tiple online social networks. IEEE Access, 5:13118–

13130.

Bier, C. and Prior, J. (2014). Detection and labeling of

personal identiﬁable information in e-mails. In IFIP

International Information Security Conference, pages

351–358. Springer.

Bracamonte, V., Hidano, S., Tesfay, W. B., and Kiyomoto,

S. (2019). User study of the effectiveness of a pri-

vacy policy summarization tool. In International Con-

ference on Information Systems Security and Privacy,

pages 186–206. Springer.

Bracamonte, V., Hidano, S., Tesfay, W. B., and Kiyomoto,

S. (2020). Evaluating the effect of justiﬁcation and

conﬁdence information on user perception of a privacy

policy summarization tool. In ICISSP, pages 142–151.

Caliskan Islam, A., Walsh, J., and Greenstadt, R. (2014a).

Privacy detective: Detecting private information and

collective privacy behavior in a large social network.

In Proceedings of the 13th Workshop on Privacy in

the Electronic Society, WPES ’14, page 35–46, New

York, NY, USA. Association for Computing Machin-

ery.

Caliskan Islam, A., Walsh, J., and Greenstadt, R. (2014b).

Privacy Detective: Detecting Private Information and

Collective Privacy Behavior in a Large Social Net-

work. In Proceedings of the 13th Workshop on Pri-

vacy in the Electronic Society, WPES ’14, pages 35–

46. Association for Computing Machinery.

Castillo, S. R. M. and Chen, Z. (2016). Using Transfer

Learning to Identify Privacy Leaks in Tweets. In 2016

IEEE 2nd International Conference on Collaboration

and Internet Computing (CIC), pages 506–513.

Mao, H., Shuai, X., and Kapadia, A. (2011). Loose tweets:

an analysis of privacy leaks on twitter. In Proceedings

of the 10th annual ACM workshop on Privacy in the

electronic society, pages 1–12.

Sleeper, M., Cranshaw, J., Kelley, P. G., Ur, B., Acquisti,

A., Cranor, L. F., and Sadeh, N. (2013). ”i read my

twitter the next morning and was astonished”: A con-

versational perspective on twitter regrets. In Proceed-

ings of the SIGCHI Conference on Human Factors in

Computing Systems, CHI ’13, page 3277–3286, New

York, NY, USA. Association for Computing Machin-

ery.

Sokolova, M., El Emam, K., Rose, S., Chowdhury, S., Neri,

E., Jonker, E., and Peyton, L. (2009). Personal health

information leak prevention in heterogeneous texts.

AdaptLRTtoND ’09, page 58–69, USA. Association

for Computational Linguistics.

Tesfay, W. B., Hofmann, P., Nakamura, T., Kiyomoto, S.,

and Serna, J. (2018). I read but don’t agree: Privacy

policy benchmarking using machine learning and the

eu gdpr. In Companion Proceedings of the The Web

Conference 2018, pages 163–166.

Tesfay, W. B., Serna, J., and Pape, S. (2016). Challenges

in detecting privacy revealing information in unstruc-

tured text. In PrivOn@ ISWC.

Tesfay, W. B., Serna, J., and Rannenberg, K. (2019). Priva-

cybot: Detecting privacy sensitive information in un-

structured texts. In 2019 Sixth International Confer-

ence on Social Networks Analysis, Management and

Security (SNAMS), pages 53–60.

Vishwamitra, N., Knijnenburg, B., Hu, H., Kelly Caine,

Y. P., et al. (2017). Blur vs. block: Investigating

the effectiveness of privacy-enhancing obfuscation for

images. In Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition Workshops,

pages 39–47.

Wang, Y., Norcie, G., Komanduri, S., Acquisti, A., Leon,

P. G., and Cranor, L. F. (2011). ”i regretted the minute

i pressed share”: A qualitative study of regrets on

facebook. In Proceedings of the Seventh Symposium

on Usable Privacy and Security, SOUPS ’11, New

York, NY, USA. Association for Computing Machin-

ery.

Wobbrock, J. O., Findlater, L., Gergle, D., and Higgins,

J. J. (2011). The aligned rank transform for nonpara-

metric factorial analyses using only anova procedures.

In Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems, CHI ’11, pages 143–

146. Association for Computing Machinery.

ICISSP 2021 - 7th International Conference on Information Systems Security and Privacy

634