This paper describes EPFL privacy protection tool for the MediaEval 2014 Visual Privacy task. The goal of the task is to obscure faces, body silhouettes, and personal items of people in the provided surveillance clips to preserve their personal privacy. The EPFL privacy protection tool mainly relies on two privacy protection filters: a warping-based reversible filter to obscure features with low visual details (body silhouettes) by distorting them with randomized warping and morphing-based reversible filter to obscure features with high visual details (faces and personal items) by 'replacing' them with a graphical representation. The aim of this tool is to achieve an acceptable balance between privacy (how well the privacy is protected) and intelligibility (how well the surveillance task can still be performed), as well as, privacy and pleasantness (how pleasant is the protection). The results of three types of subjective evaluations, via crowdsourcing, practitioners, and stakeholders, provided by the organizers of the task demonstrates that EPFL privacy protection tool achieves a great overall balance between privacy, intelligibility, and pleasantness, while being secure and reversible in the same time.
Recent adoption of digital video surveillance systems, especially
in public spaces and communities, has significantly increased the
concern for intrusion into individual privacy. New sensing
technologies, such as ultra high definition, high dynamic range, or video
capturing with mini-drones, threaten to eradicate boundaries of
private space even more. As a possible solution, many privacy
protection tools have been proposed for preserving privacy, ranging from
simple methods such as masking blurring, pixelization, or
masking to more advanced methods satisfying the following desirable
practical properties: reversibility, robustness, and security. The
advanced methods can be divided into several categories:
encryptionbased [
Despite wide availability of visual privacy protection tools, with
an exception of some work [
In the submission to MediaEval 2014 Privacy task, EPFL aimed
to address both tradeoffs and separately obscure two types of
visual features. Therefore, the proposed privacy protection tool
consists of two privacy protection filters: a warping-based filter [
Organizers of the task provided video dataset [
The best privacy preserving filter would be a blacked out camera with no video feed, but, in such case, there would be no surveillance possible and intelligibility would be zero. Therefore, a usable privacy protection filter should have a balance between privacy and intelligibility. Similarly, an encryption or scrambling based privacy filters could lead to high privacy but can be annoying or even scary, resulting in very low pleasantness. Another important practical requirement is the secure reversibility of the privacy protection tool, so that the protection can be undone in secure way (only if one has a secret key) to restore the original video in case police or court would require it.
To achieve the balance between privacy, intelligibility, and
pleasantness, as well as to provide reversible protection, the proposed
privacy protection tool adopted a two-stage approach: (i) reversible
warping filter [
Warping filter makes the details of the visible object unrecognizable (i.e., privacy is increased), but, by controlling the strength of the filter, the overall general shape of the object can be preserved, so it would still be possible to understand what is going on in the surveillance scene (i.e., intelligibility is not decreased). Morphing filter replaces faces and personal items with the graphical representation, e.g., ‘smiley face’ instead of the original face, which effectively removes all personal details, i.e., privacy is increased, with the aim to keep both intelligibility and pleasantness/appropriateness high.
The organizers provided the results from three subjective evaluations: crowdsourcing-based evaluation in Stream 1, evaluation by stakeholders and surveillance experts in Stream 2, and evaluation by practitioners and data protection experts in Stream 3. The corresponding results of the EPFL privacy protection tool are summarized in the Table 1 and compared against the average of the total 8 submissions to the Privacy Task of MediaEval 2014.
From the table, it can be noted that across all evaluations, the tool demonstrates higher than average level of privacy but underperforms in terms of intelligibility and pleasantness. In crowdsourcing evaluation, the performance of the tool is nearer to average compared to other two evaluations. It means that the tool would be more suitable for the scenarios where the observers are naïve subjects as it is in the case of crowdsourcing. The low intelligibility score can be compensated by the fact that the tool is reversible and original video can be securely restored, which would allow the detailed examination of the video data if necessary. Low pleasantness value is probably due to the choice of graphical representations for faces and personal items (see Figure 1), which subjects did not like. A more appropriate and use case oriented choice of such representation may improve the pleasantness of the visual protection. 4.
EPFL privacy protection tool combines warping and morphing privacy protection filters and achieves an acceptable balance between privacy, intelligibility, and pleasantness, providing, in the same time, ability to securely restore the original content if necessary. In a practical scenario, a better fitting graphical representations of the faces and personal items can be selected.
This work was conducted in the framework of the EC funded Network of Excellence VideoSense. 5.