Data privacy is a very critical topic in the present digital era. The data relating to a person, such as a name, age, bank details etc., require a high level of privacy. These kinds of data can be in different forms, like documents, images, sound clips, or videos.

Developments in data processing technologies have provided easier and cheaper means of processing large volumes of data more promptly. Visual data processing became more significant over the years for surveillance, mass data processing and ambient assisted systems. These developments led to the rise in data privacy risk, where image protection systems became a key player.

What is Image Data Privacy?

A balanced image protection technique can be characterized by:

• Privacy of data: Privacy from human observers or recognition systems
• Clarity of data: Any anomaly can be detected without investigating sensitive data
• Reversibility of data: The recorded data should be reconstructed with accuracy if needed
• Security: Robustness of the protection program module

Depending on the working principle of image protection, the techniques can be classified into different categories. Prominent among them are editing, face regions and false colour methods.

Editing Methods:

There are various strategies under the editing method like blurring, encryption, black box, pixelation, masking, and scrambling. Blurring overlays a Gaussian function to the image. In the blurring technique, certain operations utilize the neighbouring pixels of the sensitive area in the image to estimate pixel values of that area. In this way, blurring can cover the sensitive zone and the delicate data. For example, in Google Street View applications, the blurring technique is used to conceal data like tags and faces.

In the black-box technique, once sensitive data in the images like faces and tags are identified, the area is replaced with shapes like rectangles, ovals, etc., of solid colours like black or white.

Pixelation is another technique that helps reduce the resolution of the sensitive area in an image. Here, a particular area containing sensitive information is divided into specific small pixel units (commonly as 8×8). Then, the pixel value of these units is averaged and applied as a whole. In this way, the resolution of the sensitive area is diminished to a great extent and cannot be visible clearly. This method is generally used for the requirements like censoring news and documentary programs to protect the identity of victims, suspects, witnesses, or evidence.

Other strategies like masking utilize a strong tone for covering delicate parts once the sensitive area of an image is detected.

Editing methods are simple, but it does not separate sensitive and non-sensitive content, such that it manipulates both contents. Reproducing the original content, which is edited, is almost impossible. Although with pixelation and blurring strategies, it is achievable to an extent.

Face Regions Manipulations:

There are several methods used for face region manipulations of visual data. One of these methods uses Eigenvectors. Here, the sensitive zone is reconstructed by applying a few numbers of Eigenvectors. It makes the facial features indistinguishable and hard to recognize. Several algorithms are taken for this technique, from which one suitable algorithm is chosen to meet the requirement of security and privacy. The strategy can deliver unnatural pictures but still fit for masking some facial features.

Morphing and warping are other methods used for facial modifications. Morphing uses interpolation to map one source face to a different target face. The key points in the facial image and the intensity are changed to map the target face. In warping, randomness is added to the alignment of the essential points in the facial features in the image. Warping also follows interpolation.

These methods can keep the overall appearance and details like face and gender. The algorithms to detect objects or persons will be very weak for unclear images, and hence this method is uncommon for real-time applications.

False Color:

False-colour is another method that uses an RGB to 8-bit greyscale lookup table system. This method does not need to detect the region of interest (ROI) and therefore is an improvement over object detection algorithms. The mapping into greyscale would usually not be a perfect one to one mapping, and hence, reproduction of original data with colour information will be difficult.

Conclusion

There are different techniques for enhancing privacy for images and follow two steps toward implementing measures for protecting privacy. The first is to detect the sensitive information in an image. The next step is the manipulation of the area of interest.

Augmented Privacy for Images in iEDPS

iEDPS (Infosys Enterprise Data Privacy suite) is a data privacy protection product with support for various structured data like relational databases and files along with unstructured data like log files, word documents, images etc. It also supports several cloud platforms like Salesforce, MSDynamics, etc. iEDPS has different features like data masking, data subsetting, data generation and data discovery.

When adding privacy protection to images, iEDPS performs a sensitivity analysis on the image using various discovery configurations and identifies the sensitive region within. It is applicable for standard images like cheques, invoices, etc. For other types of images, iEDPS provides different modification techniques based on the principles discussed above.

Author: Aswathi Valsan