PaperlayerView exact revision on arXiv

Source-linked AI summary

KidsTube: Detection, Characterization and Analysis of Child Unsafe Content & Promoters on YouTube

Rishabh Kaushal, Srishty Saha, Payal Bajaj, Ponnurangam Kumaraguru

arXiv:1608.05966v1cs.SI

TL;DR

The paper asks how to detect unsafe content and its promoters around cartoon videos watched by children. It combines feature-based classification and a CNN on video frames with behavioral characterization and engagement-network analysis. It reports close communities between unsafe promoters and other users, alongside limited engagement connections among commenters and uploaders.

  • Problem

    The paper seeks to identify unsafe uploaders and commenters associated with seed cartoon videos, addressing concerns about children encountering inappropriate content on YouTube.

    Source: §I · p. 1, A. Problem Statement · p. 3

  • Method

    The study combines supervised classification using video-, user-, and comment-level features with a CNN video-frame approach, then characterizes promoters and analyzes engagement networks.

    Source: §I · pp. 1–2, A. Detecting Unsafe Content Uploaders · pp. 3, 5

  • Results

    Unsafe promoters form close communities with other users, while only 24 of 275 uploaders liked other uploaders’ videos and 604 of 19,098 commenters were connected through likes.

    Source: 1) Likes Network Community · p. 7, §VII · p. 8

  • Takeaways & Limitations

    The study concludes that child unsafe content and its promoters can be detected and that such promoters form close communities with other users.

    Source: §VII · p. 8

Abstract

from arXiv · show

YouTube draws large number of users who contribute actively by uploading videos or commenting on existing videos. However, being a crowd sourced and large content pushed onto it, there is limited control over the content. This makes malicious users push content (videos and comments) which is inappropriate (unsafe), particularly when such content is placed around cartoon videos which are typically watched by kids. In this paper, we focus on presence of unsafe content for children and users who promote it. For detection of child unsafe content and its promoters, we perform two approaches, one based on supervised classification which uses an extensive set of video-level, user-level and comment-level features and another based Convolutional Neural Network using video frames. Detection accuracy of 85.7% is achieved which can be leveraged to build a system to provide a safe YouTube experience for kids. Through detailed characterization studies, we are able to successfully conclude that unsafe content promoters are less popular and engage less as compared with other users. Finally, using a network of unsafe content promoters and other users based on their engagements (likes, subscription and playlist addition) and other factors, we find that unsafe content is present very close to safe content and unsafe content promoters form very close knit communities with other users, thereby further increasing the likelihood of a child getting getting exposed to unsafe content.

I. INTRODUCTION

The paper studies unsafe videos and comments around cartoon content watched by children, and introduces detection, characterization, and network analyses of their promoters. It distinguishes this focus from general YouTube spam and promotion research.

I. INTRODUCTION

  • The study focuses on cartoon videos watched by children, where unsafe content may appear alongside widely viewed safe cartoons.The introduction illustrates safe cartoon videos, unsafe cartoon videos, and unsafe comments on otherwise safe videos.Source: §I · p. 1, Figure 1 · p. 2
  • Unsafe content appears as indecent videos or inappropriate comments, including comments that make an otherwise safe video page inappropriate.The paper defines indecent videos as containing graphic nudity or abusive or inappropriate dialogue.Source: §I · p. 1
  • Prior reports raised concerns about children reaching adult content, while YouTube Kids relied on viewer flagging that could be manipulated.Consumer groups reported inappropriate content in the app to the FTC.Source: §I · p. 1
  • The paper presents supervised classification using video-, user-, and comment-level features, alongside a CNN approach using video frames.It also compares unsafe promoters with other users and analyzes their engagement-based communities.Source: §I · pp. 1–2

A. Detection of Inappropriate Content on YouTube

The paper frames detection, characterization, and community discovery for unsafe content promoters among uploaders and commenters. It builds on prior inappropriate-content and malicious-user studies while focusing on unsafe promotion around cartoon videos.

A. Detection of Inappropriate Content on YouTube

B. Detection of malicious user behaviour

  • Earlier YouTube research examined spammers, promoters, fraudulent promotion, comment spam, and hate or extremist video retrieval [14] [13].This paper instead studies inappropriate words in comments and relationships between commenters and uploaders.Source: B. Detection of malicious user behaviour · p. 2

A. Problem Statement

  • The detection problem is to identify unsafe uploaders and commenters from seed cartoon videos, their uploaders, and their commenters.The paper separately defines characterization of safe versus unsafe uploaders and community detection within and across uploader and commenter groups.Source: A. Problem Statement · p. 3

B. Data Collection

The dataset begins with cartoon-keyword search results and expands through user engagements. Ground truth labels videos through manual inspection and identifies comments containing words from a bad-word list.

B. Data Collection

  • The seed dataset contains 408 randomly selected cartoon videos, 275 uploaders, and 21,268 comments from 19,099 unique commenters.The search used the top 20 popular cartoon keywords; up to 100 latest comments were collected per video.Source: B. Data Collection · p. 3
  • Engagements expanded the dataset to videos liked, uploaders subscribed to, and videos playlisted by uploaders and commenters.The collection includes 1,178 videos liked by uploaders and 123,390 liked by commenters, as well as subscription and playlist data.Source: B. Data Collection · p. 3, Table II · p. 3

C. Building Ground Truth

IV. DETECTION OF UNSAFE CONTENT PROMOTERS

The paper detects unsafe uploaders with a classifier built from video-, user-, and comment-level features, and also tests a CNN approach on video frames. The CNN approach achieved lower accuracy and was computationally expensive.

A. Detecting Unsafe Content Uploaders

  • The classifier approach marks uploaders safe or unsafe using video-level, user-level, and comment-level features.The feature sets capture video responses and metadata, uploader profile attributes, and responses to comments.Source: A. Detecting Unsafe Content Uploaders · pp. 3–4, Table III · p. 4
  • Random forest performed best among the conventional classifiers for predicting video safety on the seed dataset using an 80:20 train-test split.The resulting video classifier was applied to up to 50 latest videos per unique seed-dataset uploader.Source: A. Detecting Unsafe Content Uploaders · p. 4, Table IV · p. 4
  • Uploaders are categorized as extreme, high, or moderate promoters according to the degree of unsafe content in their past uploads.The ratio of indecent videos to total videos informs thresholds for classifying uploaders as safe or unsafe.Source: A. Detecting Unsafe Content Uploaders · pp. 4–5, Figure 2 · p. 4
  • 74% accuracy was achieved by the CNN video-frame approach, which was less accurate than the classifier approach and computationally expensive.It detected dissimilar frames after processing abrupt and gradual visual transitions, with post-processing for movement and flash-light outliers.Source: A. Detecting Unsafe Content Uploaders · p. 5

B. Detecting Unsafe Content Commenters

The study labels a user an unsafe commenter if they posted at least one comment containing indecent words.

B. Detecting Unsafe Content Commenters

  • At least one comment containing indecent words qualifies a user as an unsafe commenter.Of 21,268 comments, 1,814 contained bad words and were posted by 1,755 unique users, all marked unsafe.Source: B. Detecting Unsafe Content Commenters · p. 5

V. UPLOADER CHARACTERIZATION STUDY

The characterization study compares unsafe and safe content uploaders by their popularity and engagement to identify distinguishing features.

V. UPLOADER CHARACTERIZATION STUDY

  • The study compares unsafe and safe uploaders’ popularity and engagement to identify distinguishing features.It examines engagement received on videos uploaded by the identified unsafe uploaders.Source: §V · p. 5

A. Popularity Factors

The study characterizes uploader popularity and engagement, then uses community analysis to examine connections between safe and unsafe content promoters.

A. Popularity Factors

  • Unsafe uploaders have fewer subscribers, uploaded fewer videos, received fewer views, and had smaller Google+ user networks than safe uploaders.The authors suggest the lower viewership could be due to unsafe uploaders posting fewer videos.Source: A. Popularity Factors · p. 5

B. Engagement Factors

  • Safe content uploaders received more comments, likes, and dislikes on their videos than unsafe content promoters.The authors interpret this as users, typically children, engaging more with those spreading safe content.Source: B. Engagement Factors · p. 5

VI. COMMUNITY DETECTION

  • Community detection examines whether safe and unsafe promoters are closely connected, which could help explain movement from safe-content promotion toward unsafe promoters.The analysis includes both uploaders and commenters.Source: §VI · p. 5

A. Video Network Analysis

The video-network analysis links seed videos through YouTube’s related-video suggestions and examines transitions among safe and unsafe videos.

A. Video Network Analysis

  • The graph links a video to suggested related videos that are also in the 408-video seed set, yielding 262 nodes and 630 edges.The analysis takes the top 10 related videos for each seed video.Source: A. Video Network Analysis · p. 6
  • The graph contains 31 communities with Louvain Modularity 0.807, and Figure 5 depicts four transition types between decent and indecent videos.Table V reports the distribution of the 630 video-video and uploader-uploader transitions.Source: A. Video Network Analysis · p. 6, Figure 5 · p. 6, Table V · p. 7
  • 9 safe-to-unsafe transitions contrasted with 117 unsafe-to-unsafe transitions in the video-video graph; the paper also reports 14 safe-to-unsafe uploader transitions.The authors infer that a user entering unsafe video space may remain there through video suggestions.Source: A. Video Network Analysis · pp. 6–7, Table V · p. 7

B. Uploader-Commenter Network Analysis

The uploader-commenter graph links users through comments, while the behavioral analysis examines likes, subscriptions, and playlist additions to map further connections.

B. Uploader-Commenter Network Analysis

  • 733 edges link unsafe commenters to safe uploaders, showing how comments can undermine otherwise safe uploader behavior.The graph contains 1,942 commenter and uploader nodes and 2,373 edges; its 147 communities have a Louvain Modularity value of 0.973.Source: B. Uploader-Commenter Network Analysis · p. 7, Table VI · p. 7

C. Uploader-Commenter Behavioural Analysis

  • Likes, subscriptions, and playlist additions define the behavioral engagements used to examine communities among safe and unsafe uploaders and commenters.Relationships are established by reverse lookup against the seed dataset because the YouTube API does not directly provide lists of users who liked a video or subscribed to a user.Source: C. Uploader-Commenter Behavioural… · p. 7
  • The behavioral analysis uses 275 seed uploaders and 19,099 seed commenters, with results summarized across three engagement types.The procedures for constructing these behavioral relationships are omitted for space considerations.Source: C. Uploader-Commenter Behavioural… · p. 7

1) Likes Network Community:

Likes, subscriptions, and playlists reveal different patterns of connection among uploaders and commenters, including shared targets and differences in whose content they engage with.

1) Likes Network Community:

  • Only 24 of 275 uploaders like other uploaders, versus 604 of 19,098 commenters connected through likes; both groups like 262 common other users.The shared users suggest similar interests between uploaders and commenters.Source: 1) Likes Network Community · p. 7

2) Subscriber Network Community:

  • Around 10% of uploaders connect through subscriptions, versus 617 of 19,098 commenters; commenters subscribe to more uploaders, while the reverse is rare.There are no self-subscribers in this analysis.Source: 2) Subscriber Network Community · p. 7

3) Playlist Network Community:

  • 757 of 1,068 uploader playlist connections are self-additions; commenters include only 2,148 of 30,148 commenter-uploaded videos, and both groups add videos from 67 common other users.Most videos in commenters’ playlists are from other users.Source: 3) Playlist Network Community · pp. 7–8

4) Combined Network Community:

The combined engagement graph shows close communities of safe and unsafe content promoters; the paper concludes that unsafe promoters also form close communities with other users.

4) Combined Network Community:

  • The combined graph’s closely knit safe and unsafe communities reinforce concern that a child watching a safe promoter could move into unsafe promoters’ space.Figure 6 depicts unsafe promoters in red and safe promoters in blue.Source: 4) Combined Network Community · p. 8, Figure 6 · p. 8
  • The combined graph includes uploaders and commenters connected through likes, subscriptions, or playlist relationships.The authors describe the proximity of unsafe promoters to safe content promoters as increasing the likelihood of child exposure.Source: 4) Combined Network Community · p. 8

VII. CONCLUSION

  • The authors report detecting child-unsafe content and its promoters, characterizing promoter popularity and engagement, and finding close communities with other users.Source: §VII · p. 8

Related papers

Loading 1608.05966v1…