This paper introduces a new privacy lter adopted in the context of the DPT (Drone Protect Task) at MediaEval Benchmark 2015. Our proposed lter protects privacy by visually replacing sensitive RofI (Regions of Interest) by its shapes. A combination of steganography and scrambling is used in order to make this lter. Once the scrambling is applied on the pixels of the RofI, its MSB (Most Signi cant Bit) are hidden in the LSB (Least Signi cant Bit) of a cover image. Our lter ful ls four criteria de ned by DPT: near-lossless reversibility, intelligibility, appropriateness and anonymization. We benchmarked the lter on the last three criteria and we get good results: 40 % for intelligibility and appropriateness, and 60 % for anonymization.
Due to the growing of video surveillance systems and the signi cant improvement of automatic recognition tools, privacy protection techniques became a necessity. Moreover, these systems bene t from image sensors progress (e.g. people are recognized far away from the camera).
Here examples of already existing systems protecting privacy: pixelization, blurring or black masking with FacePixelizer 1 on Google plus, ObscuraCam 2 on Android, and also scrambling in JPEG compression with Scrambling JPEG tool 3. Besides, people are working on methods to hide identity such as morphing [5], warping [5] and scrambling [6], but they are complex and the degradation they apply on the images prevent any usage for security purpose (lack of intelligibility).
The use-case scenario designed for the challenge was Car Park Security. The goal was the creation of privacy ltering solutions for drone videos related to public safety. They are evaluated by following four criteria: i) protection of privacy, ii) intelligibility for the visual quality in order to recognize events on the result (i.e. people walking, running, ghting, stealing...), iii) appropriateness to see if the result is good looking, iv) possibility to reverse to come back to the original im1http://www.facepixelizer.com/ 2https://guardianproject.info/apps/obscuracam/ 3http://ltslinux18.ep .ch/scramble/ 2.
RofI (e.g. people, vehicles or accessories bounding boxes)
are previously annotated in the database [2] used for DPT [
To hide information, an XOR is computed between the six MSBs of the RofI and the random numbers generated with a PRG (pseudorandom generator) controlled by a seed, as expressed in the equation 1.
XORImg(i) = Rof I(i)
RandN ums(i); 8i (1) with i the bit position and each bit 2 f0, 1g.
In parallel, cover images are computed to replace RofI and keep the possibility to recognize events. An edge detector and a Kmeans clustering [3] (limited to two clusters, similar to a binarization) are applied in the RGB space of the RofI containing people. An AND is computed between the edges of the RofI and the resulting clusters of the images on each pixel, by multiplying them as shown in the equation 2.
CoverImg = EdgeImg: KmeansClustering; (2) with .* the Element-by-element multiplication.
The convex hull image from the binary cover image for people is generated in order to become the RofI containing people. RofI containing car or accessories use(s) only Kmeans clustering as cover image.
Next, the 2 MSBs of the cover image, where the pixels intensity is either 192 or 0, are inserted in the 2 MSBs of the resulting image. Finally, the 6-bit of the XOR image, where pixels intensity is between 0 and 63, are integrated in the LSB of the resulting image as shown in the equation 3. Therefore, only cover images are visible by viewers in order to recognize events.
P rotectedImg = 5 7 X XORImg(i) 2i+X CoverImg(i) 2i; i=0 i=6 (3)
Figure 1 illustrates the work ow of the proposed method and Figure 2 shows an example of an entire privacy image. ePassphrase eSeed
HHH
RofI
HHH
XOR 2-Bit = 2 MSB
6-Bit = 6 LSB
To recover the original RofI, the inverse process is applied as shown in the equation 4. Two LSBs are removed from the original RofI, thus, a maximum error rate of 3 may be produced between an original pixel and a recovered pixel. This error implies no impact for human vision and is negligible for machines.
7 Recovered = X(P rotectedImg(i 2) RandN ums(i)) 2i i=2 (4) 2.1
One pixel is considered with 8-bit from MSB to LSB.
Original pixel b7 b6 b5 b4 b3 b2 b1 b0 For each pixel of the RofI, only the MSB bits between 2 and 7, are preserved. An XOR is computed between the MSB of the original pixel and a random number. The result is denoted b0.
XORpixel, b0 b07 b06 b05 b04 b03 b02 X X The bits of b0 are shifted in the six LSBs.
XORpixel, b0 X X b07 b06 b05 b04 b03 b02 Two MSBs of a white and black pixel are represented by e6 = 1 and e7 = 1 for the former, and e6 = 0 and e7 = 0 for the latter. Finally, the two MSBs of e are added with the six LSBs of b0. The result is denoted, protected pixel. To recover the original pixel, an XOR is computed between the same random number than previously, and the six LSBs of the protected pixel.
Recovered pixel b7 b6 b5 b4 b3 b2 X X The process is mostly reversible because the two LSBs, denoted b0 and b1 are lost. 3.
We tested our proposed lter on di erent video sequences
from DronesProtect dataset [2]. The guidelines of the
MediaEvel 2015 DroneProtect Tasks [
The aim of the challenge is to nd a trade-o between privacy and visual quality of the protected image. Indeed, the higher is the protection, the lower is the level of information (see intelligibility and appropriateness).
Two human evaluator groups are selected. In the rst group, people come from surveillance security domain (R & D), and in the second group they come from any other domain (Naive).
In Table 1, we report the average results of our lter. We obtained positive feedbacks from the jury and especially for the privacy protection. Indeed, according to the results 60 % of privacy is well protected. However, we got 40 % for intelligibility and appropriateness; this shows a lack in our lter for recognizing events properly. This can be explained because the edges detector makes mistakes and also colors of RofI are turned to black and white. It is planned as future work to improve the edges detection method with a new design for the cover image, in order to be better tailored to release more information and having a better event recognition. [2] M. Bonetto, P. Korshunov, G. Ramponi, and T. Ebrahimi. Privacy in mini-drone based video surveillance. In Workshop on De-identi cation for privacy protection in multimedia, number
EPFL-CONF-206109, 2015. [3] H. Fradi, Y. Yan, and J.-L. Dugelay. Privacy protection lter using shape and color cues. In MediaEval 2014 Workshop, Barcelona, Spain, October, 2014. [4] J. Hu and F. Han. A pixel-based scrambling scheme for digital medical images protection. Journal of Network and Computer Applications, 32(4):788{794, 2009. [5] P. Korshunov and T. Ebrahimi. Towards optimal distortion-based visual privacy lters. In IEEE International Conference on Image Processing, number EPFL-CONF-197087, 2014. [6] A. Melle and J.-L. Dugelay. Scrambling faces for privacy protection using background self-similarities. In Image Processing (ICIP), 2014 IEEE International Conference on, pages 6046{6050. IEEE, 2014.