PhD proposal: Suman Jana, December 2, 2013, 12:00 PM CST, GDC 5.816

Contact Name: 
Lydia Griffith
Dec 2, 2013 12:00pm - 2:00pm

Ph.D. Oral Proposal: Suman Jana

Date: December 2nd, 2013
Time: 12:00 PM CST
Location: GDC 5.816

Supervising professor: Vitaly Shmatikov

Title: Security and Privacy in Perceptual Computing


Perceptual, "context-aware" applications that observe their environment and
interact with users via cameras and other sensors are becoming ubiquitous on
personal computers, mobile phones, gaming platforms, household robots, and
augmented-reality devices. Such applications present several new security and
privacy challenges.

This thesis shows that the existing security and privacy protection mechanisms
in perceptual computing platforms are inadequate and explores two new approaches
for building new, privacy-preserving platforms. We find that inadequate system
support forces existing perceptual computing platforms like augmented reality
browsers to implement ad hoc mechanisms that contain numerous inherent security
and privacy flaws. We demonstrate that such flaws can be exploited by untrusted
perceptual web content for cookie theft, cross-site scripting attacks against
any Web content, bypassing of normal access control for device resources (e.g.,
the on-board camera), clickjacking etc.

To avoid such problems, we design and implement two complementary approaches for
building privacy-preserving perceptual computing platforms. First, we build
DARKLY, a privacy protection system that deploys three different privacy
protection mechanisms to achieve this goal - access control, algorithmic privacy
transforms, and user audit. DARKLY is integrated with OpenCV, a popular computer
vision library used by most perceptual applications to access visual inputs.
Next, we build a new OS abstraction for perceptual applications - the
recognizer. Instead of exposing privacy-sensitive raw sensor data to the
applications directly, a recognizer detects higher-level objects (e.g., a face)
from raw sensor data and only exposes these objects to the application. We show
how such recognizers can be used for creating a fine-grained permission system
to support development of least privilege perceptual applications.