Investigation, Detection and Prevention of Online Child Sexual Abuse Materials: A Comprehensive Survey

[rezaBarzgar]

Survey overview

This comprehensive survey focused on papers published between 2018 and 2022 with the purpose of online Child Sexual Abuse Materials (CSAM). Voung et al.^1 claim that they are the first to survey state-of-the-art online CSAM papers and research. They reviewed 43 Papers

Reviewed Papers

---

Confrence	Number of literature
Springer	13
Elsevier	12
IEEE	7
ACM	4
Others	7

Motivation

---

I think this survey proposed legit motivations for the essence of online child abuse detection. Their first motivation is long-term physical and psychological effects. Due to the fact that most of the sexual abuses are unreported (about 69%), they mentioned online sexual abuse detection is necessary as their second motivation.^2.

Literatures Hierarchy

based on the type of dataset

Dataset		open source
self-build	21	1
third-party	14	10

self-build: they crawled social media and websites and they did not share their data, except for one. Unfortunately, the data is image and video. Therefore, we cannot use it.^3 third-party: Four of the papers used PAN2012 as data, and others used the following datasets:

• https://github.com/shaheerakr/roman-urdu-abusive-comment-detector
• https://www.ndacan.acf.hhs.gov/
• https://www.icpsr.umich.edu/web/HMCA/studies/6647
• https://www.cdc.gov/violenceprevention/childabuseandneglect/vacs/
• https://www.nist.gov/itl/ssd/software-quality-group

---

Based on the type of task addressed

task addressed
Text	24
Image	9
Video	3

Text: Those who used text as their input data are in two domains:

• computer science: they proposed methods to detect abusive comments, discover the behavior of predators, and identify predatory conversations
• social sciences: they mostly focused on the effects of CASM and its relation with other features such as culture, race, and ethnicity.

Image: Those who used child sexual abuse images as their input data worked on two domains:

• detection of physical signs to help hospital workers investigate CSA.
• deleting CSA images from social media

Video: Those who used video as their input data worked on processing videos to prevent sharing of CSA images and videos on online sharing media.

---

Based on the type of technique

technique
Statistic	14
NLP	4
Machine learning / Deep learning	16

NLP: In ^4, the list of keywords about emotions and sentiments posted on social media was built. Then, the authors analyzed the Twitter tweets and Facebook pages of some potential people in India based on the keyword list. This can be interesting.

ML / DL: Bours and Kulsrud^5 applied CNN and Naıve Bayes based on features of message, author, and conversation. It has 20 citations. I must read it.

---

Survey Link

https://ieeexplore.ieee.org/abstract/document/10013853/

[rezaBarzgar]

@hosseinfani Please take a look at this literature review.

[hosseinfani]

@rezaBarzgar Thank you. Nice summary. Also very nice catch.