Flasher Detection Project in Online and Mobile Video Chat


A key area of  CyberSafety   research  that  we  are  exploring  concerns  making  the  Internet  safer for users  of  popular random video chat applications  like  Chatroulette  and  Omegle.  In  such  applications,  one  of  the  major  problems  experienced  by  users  is  that  of  flasher misbehavior. Such  misbehavior  can  be  distasteful  and  potentially  illegal  if  minors  are  present  on  the site. Our research consists of the following efforts:
  • Safevchat (published in WWW 2011): our original first paper studying how to automatically and accurately detect flasher behavior on the popular online random video chat site of Chatroulette. We found that there was a high correlation between the lack of a face and flasher behavior, and used this as a basis for our fusion classifier. See safevchat.org for more information.
  • Fine-grained Cascaded Classification (FGC) (published in WSDM 2012): we addressed how to scale the machine learning based classifier for flasher detection so that fewer server resources were consumed. By cascading the classifier, we were able to reduce the number of servers needed by a factor of about a third.
  • Emerald (published in KDD 2012): proposing an alternative method to scale the server-based classifier using a rule-based pre-classifer followed by linear regression (sparingly invoked) in a multi-phase process.
  • mvchat (published in UbiComp 2013): we extended our work from the previous three papers to address mobile random video chat. We built a mobile video client that was compatible with the Omegle random video chat Web site. Over 5000 users downloaded our client. Our interest was in understanding how the mobile scenario differs from the online scenario. Our user studies found that the correlation between the lack of a face and flasher behavior seen in online video chat was much weaker in mobile video chat, due to the wider diversity of images from the mobile camera. See also below for more correlation findings.
  • Multi-modal Fusion for Flasher Detection in a Mobile Video Chat Application (published in Mobiquitous 2014): based on the user studies and correlations we found in the UbiComp 2013 paper, we determined that using multi-modal mobile sensor data and temporal data can substantially improve the accuracy of the fusion classifier, compensating for the loss in accuracy due to the weaker correlation between facial absence and flasher behavior.

Flasher Detection Analysis on Random Web Video Chat Application

Our previous studies about detecting flasher on random web video chat are shown in safevchat.org

Flasher Detection Analysis on Random Mobile Video Chat Application

Online video chat services such as Chatroulette and Omegle randomly match users in video chat sessions and have become increasingly popular, with tens of thousands of users online at anytime during a day. Our interest is in examining user behavior in the growing domain of mobile video, and in particular how users behave in such video chat services as they are extended onto mobile clients. To date, over four thousand people have downloaded and used our Android-based mobile client, which was developed to be compatible with an existing video chat service. Our project provides a first-ever detailed large scale study of mobile user behavior in a random video chat service over a three week period. This study identifies major characteristics such as mobile user session durations, time of use, demographic distribution and the large number of brief sessions that users click through to find good matches. Through content analysis of video and audio, as well as analysis of texting and clicking behavior, we discover key correlations among these characteristics, e.g., normal mobile users are highly correlated with using the front camera and with the presence of a face, whereas misbehaving mobile users (flasher) have a high negative correlation with the presence of a face.


                                                                             A screenshot of the MVChat application.

We label the snapshots in long sessions and those sessions (by major voting from image labels) into multiple taxonomies. And then we analyze the correlations between different taxonomies and try to  identify behavioral characteristics which are likely or unlikely to occur together. Below  summarizes the key results of our taxonomy correlation analysis, including strong positive (negative) correlations, some (surprising) non-correlations, and one-directional association rules with high confidence values on both image(L) and session(R) level. Our interesting and important findings include:  Normal users are highly correlated with using the front camera and showing their faces, whereas misbehaving users tend to hide their faces – which suggests the exploration of camera position and face detection for distinguishing normal users from misbehaving ones; Users with a large enough fraction of sustained sessions are disproportionately female, but surprisingly females were just as likely to misbehave as males.

Figure (d) Taxonomy correlation analysis: (L) image-based graph and (R) session-based graph. Highlighted in the graphs are strong positive correlations (both all_conf and cosine are > 0.85), strong negative correlations (both all_conf and cosine are < 0.1), some non-correlations (X^2 is close to 0 and lift is close to 1), and association rules with confidence > 0.8.

To learn more interesting findings in our project, you can go to here for details.

Comments