Obfuscated Volume Rendering

Jia-kai Chou and Chuan-Kai Yang

Computer Graphics & Multimedia Lab., NTUST


 

Abstract

Analyzing and processing various data types in a privacy-preserving perspective has been researched in many disciplines; however, such an issue draws very limited attention in the research field of scientific visualization. We wondered if it is possible to delegate the rendering of a volume data set to a remote server(s) while still being able to preserve its privacy to certain extent. This paper presents a block-based volume data transformation algorithm that obfuscates a volume data set so as to reduce the user’s privacy concern when the volume data set is to be uploaded to a remote server. In addition, a privacy-aware transfer function adjustment is proposed so that not only the privacy is protected during the rendering process, but also the computational loading could be leveraged to the server side as much as possible. Experimental results show that the proposed method yields visually satisfactory results compared with a normal direct volume rendering approach. Moreover, the decrease of the rendering efficiency caused by the proposed method is still controlledwithin an acceptable range.A case study proves that the proposed approach can be adopted in practice. This work explores the possibility of rendering a volume data set through remote server(s) while the privacy of data is still maintained.

[Paper] [Supplementary Results]

Figure 1: (a) A rendering result of the “MRBrain” data set (256 × 256 × 109 voxels) and its corresponding voxel value histogram and a user-assigned transfer function. (b) The same transfer function applied to the transformed “MRBrain” data set.

Figure 2: The overview of the flow of the remote rendering process. 1) The user-assigned transfer function is used to generate several adjusted transfer functions, which are then sent to the server side. 2) The server side renders the corresponding sub-volumes according to the given adjusted transfer functions. 3) The server side sends back the obfuscated rendering results to the client side. 4) The client side keeps only the required rendering results while discarding the rest. 5 )The client side performs the inverse permutation to set the rendering results of each sub-volume in the desired order, and then composites them altogether to get the final image.

Figure 3: A step-by-step illustration of the proposed transfer function adjustment. For a transfer function, we first remove its color information. Then, the color ranges are divided into segments according to the assigned alpha values. The segmentation boundaries and the segments with non-zero alpha values are highlighted in red. After that, the segments are linearized into pieces of straight lines. Finally, each segment with non-zero alpha values is used to generate a new adjusted transfer function with equalized slope and intercept in all segments. The corresponding segments in the original transfer function that are used for creating the adjusted transfer functions are marked in blue.

The rates of frame per second (fps) required while exploring the volume data sets are measured to compare the rendering efficiency between the proposed method and the normal volume rendering technique.

Case Study

Comparing the volume rendering results of human brains between a healthy subject (a) and two PD patients (b) and (c). In addition, each human brain is rendered by both normal direct volume rendering technique (shown on the left column) and the proposed obfuscated volume rendering approach (shown on the right column). It shows that the proposed method yields comparable rendering images with the normal rendering technique. As can be seen in a and b, it is obvious that the shape and silhouette of bilateral caudate, NAc, APu and PPu are more symmetric and visible in the health subject (a) than in PD Patient 1 (b). However, when a PD patient is with mild symptoms (PD Patient 2), the reduction of VMAT2 may not be obvious enough for one to tell whether the subject suffers from PDor not. In such a case, it may require us to extract a specific slice(s) of the data for further investigation.

Approximating the rendering of one single slice or any specifically chosen slices by rendering and compositing nearby sub-volumes. In this example, only the sub-volumes such as bilateral caudate, APu and PPu reside in are involved. It can be seen that the contralateral PPu gradually reduces in PD patient 2, while the features in the healthy subject are more symmetric and clearly visible. Moreover, we also show that the images rendered by the proposed approach (shown on the right column) are close enough to the results generated by a normal rendering method without affecting the visual assessment.