Evaluating image filtering based techniques in media space applications

Abstract

Nledia space applications that promote informrd aw~eness in an organization confront an inevitable paradox: the shared video connections between offices and rooms that promote informal awareness rdso cm rob individurds of privacy. A important open problem in this area is how to foster awareness of colleagues while minimizing the accompanying 10SSof privacy. One proposrd put forward is to filter the communicated video streams rather than broadcasting clear video. Such a scheme may facilitate awareness while helping to tieviate some aspects of the privacy loss. In tis ardcle, we describe several image filtering twhniques that provide awareness in informrd group communication applications while blurring the details of an individurd's activities, thus potentially preserving more privacy. We describe studies to quantitatively and qurditatively assess tie degrees of awareness and accuracy that these filtering tmtilques provide.

Integrating Utility into Face De-identification

“…In previous studies, it has been claimed that ad hoc de-identification methods, such as pixelation and blurring prevent humans from reliably recognizing the identity of de-identified images [7,4,8,9,5,11]. However, as our experiments demonstrate, these methods can not prevent a computer from reliably performing recognition; even at 40% de-identification the PCA algorithm achieves almost perfect recognition.…”

“…In particular, proposed methods include pixelation and blur and various other distortion filters [7,4,8,9,5,10,11]. This set of obfuscating techniques prevent the rendering of the original image, but they do not provide intuition as to whether or not privacy is maintained.…”

“…This set of obfuscating techniques prevent the rendering of the original image, but they do not provide intuition as to whether or not privacy is maintained. To test if these techniques do protect privacy, researchers investigate the degree to which these methods fool human observers of the obfuscated video [7,10,11]. Researchers ask individuals if they can determine the scene activities or identify the subjects.…”

Achieving Anonymity against Major Face Recognition Algorithms

“…the eyes) or blurring or pixelation of faces [7,17,20,36] were eventually tested. However, these methods are visually intrusive and target human and algorithmic recognition alike.…”

Model-Based Face De-Identification

“…The majority of work on images or image sequences applies simple distortion methods such as "pixelation" (image subsampling) or "blurring" (smoothing the image with e.g. a Gaussian filter with large variance) to obfuscate parts or all of the image [7,17,20,27]. The PrivacyCam architecture proposed by Senior et al [24] suppresses automatically segmented foreground objects in the scene and cryptographically secures access to the altered video stream produced by the system.…”