Presentation Title
VirtualIdentity: A Privacy-Preserving User Profiling Service
Degree Name
Master of Computer Science and Systems (MCSS)
Department
Institute of Technology
Location
Carwein Auditorium (KEY 102), UW Tacoma
Event Website
http://guides.lib.uw.edu/tactalks
Start Date
19-5-2016 6:05 PM
End Date
19-5-2016 6:10 PM
Abstract
As more users are creating their own content on the web, there is a growing interest for companies to mine this data for user profiling, i.e., to automatically guess the gender, age, personality traits, and other characteristics of their users. In addition to myriad applications in e-commerce, there is a growing interest around user profiling in digital text forensics as well. These existing personalized digital forensic applications are websites where people input their personal information for fun – for example, to get insight into how old they look, how hot they look, etc. To do so, users have to send their personal data – such as their profile pictures and their social network status updates – to the server of those websites. People who care about their privacy are understandably concerned about sending their personal information like this. Although organizations that run this kind of services usually promise not to share or sell users’ information to other parties, they can still retain and use the information they gathered. People who work in those organizations will be able to review all the information, and even if it is forbidden to trace it back to their users, it can technically still be done. In this talk we present VirtualIdentity, a privacy-preserving user profiling service that employs cryptographic protocols to infer a user’s gender, age, and personality traits from profile pictures and status updates without actually seeing the picture or the text. One way to do so would be to run the machine learning models on the client side. The obvious drawback of that approach is that it requires exposing the models – which are often considered business secrets – to the clients. Traditional cryptography does not help us here, as data encrypted with usual techniques becomes useless, it cannot be processed. Therefore, we develop novel cryptographic techniques that not only protect private information but also enable one to perform mathematical operations on encrypted data. The personal information users upload is encrypted at the client side before being sent to the server. Then, by using our novel secure multi-party computations protocols, the server can still give out the correct prediction to users. Because of the underlying cryptographic protocols, users of VirtualIdentity do not have to worry about personal information being leaked while using our personalized digital forensic services. To the best of our knowledge this is the very first platform to compute personality prediction while preserving the privacy of the user’s data and keeping the prediction model secret as well.
COinS
VirtualIdentity: A Privacy-Preserving User Profiling Service
Carwein Auditorium (KEY 102), UW Tacoma
As more users are creating their own content on the web, there is a growing interest for companies to mine this data for user profiling, i.e., to automatically guess the gender, age, personality traits, and other characteristics of their users. In addition to myriad applications in e-commerce, there is a growing interest around user profiling in digital text forensics as well. These existing personalized digital forensic applications are websites where people input their personal information for fun – for example, to get insight into how old they look, how hot they look, etc. To do so, users have to send their personal data – such as their profile pictures and their social network status updates – to the server of those websites. People who care about their privacy are understandably concerned about sending their personal information like this. Although organizations that run this kind of services usually promise not to share or sell users’ information to other parties, they can still retain and use the information they gathered. People who work in those organizations will be able to review all the information, and even if it is forbidden to trace it back to their users, it can technically still be done. In this talk we present VirtualIdentity, a privacy-preserving user profiling service that employs cryptographic protocols to infer a user’s gender, age, and personality traits from profile pictures and status updates without actually seeing the picture or the text. One way to do so would be to run the machine learning models on the client side. The obvious drawback of that approach is that it requires exposing the models – which are often considered business secrets – to the clients. Traditional cryptography does not help us here, as data encrypted with usual techniques becomes useless, it cannot be processed. Therefore, we develop novel cryptographic techniques that not only protect private information but also enable one to perform mathematical operations on encrypted data. The personal information users upload is encrypted at the client side before being sent to the server. Then, by using our novel secure multi-party computations protocols, the server can still give out the correct prediction to users. Because of the underlying cryptographic protocols, users of VirtualIdentity do not have to worry about personal information being leaked while using our personalized digital forensic services. To the best of our knowledge this is the very first platform to compute personality prediction while preserving the privacy of the user’s data and keeping the prediction model secret as well.
https://digitalcommons.tacoma.uw.edu/tactalks/2016/spring/9