Privacy Protecting Vidio Surveillance

Privacy Protecting Video Surveillance Sen-ching Samson Cheung, ECE Department, University of Kentucky Research Goals and Objectives and Background Video surveillance has become part of our daily lives. Whether we notice the cameras or not, they are mounted in countless shopping malls for deterring crimes, at toll booths for assessing tolls, and at traffic intersections for catching speeding drivers. Since the 9-11 terrorist attack, researchers have been called upon to study the use of advanced pattern recognition algorithms in video surveillance. The objective is to turn the labor intensive surveillance monitoring process into a powerful automated system needed for counter-terrorism. There are now algorithms for tracking and recognizing people across a wide-area surveillance network and algorithms for identifYing events like people gathering, dropping, and carrying various types of objects. These sophisticated algorithms, however, have much potential for being misused. An article by the American Civil Liberties Union has cited five possible ways of abusing video surveillance: 1) criminal abuse by "bad apples" in law enforcement, 2) institutional abuse by governments spying on activists, 3) abuse for personal purposes like stalking, 4) discriminatory targeting such as profiling, and 5) voyeurism [1]. The core of all these abuses is the violation of individual's privacy. The goal ofthis project is to investigate and develop novel techniques to protect privacy information in video surveillance without comprising its usefulness. In many areas of computer science such as e-commerce, networking, and data mining, researchers have begun studying algorithms that can protect individuals' privacy without compromising the benefits brought forth by the new technologies. However, until recently, there is relatively little research on privacy protection for video surveillance. The researchers at IBM have done pioneering work in this field by laying out various legal issues and technical requirements [9]. The Data Privacy Group at Carnegie Mellon University has developed a face modification algorithm to counter face recognition [8]. Other research includes the privacy-protecting data collection project at UC Irvine [12] and our own work on data hiding for protecting privacy information [15]. Despite of these efforts, we are still a long way from building a practical surveillance system that can fully protect privacy. As such, I would be much obliged if the ORAU committee can help me to jumpstart my research on this important topic. My primary research objectives are: first, to evaluate the appropriateness of the existing privacyprotecting technologies for particular surveillance applications; second, to develop novel signal processing and computer vision techniques for protecting and preserving privacy information; and third, to demonstrate the feasibility of such techniques by building a prototype system. Expected Research Outcomes and Relevance A practical privacy-protecting surveillance system shares many common elements with an ordinary surveillance system such as object tracking, face identification, database support for access control, etc. To differentiate our research with others, we will focus on our research outcomes upon two aspects unique to privacy protection: altering video to conceal private information and preserving it in a secure manner. The most common video alteration technique for privacy protection is to use a black box or large pixels (pixilation) to cover a person's face or the name of a store. Though simple to implement, these techniques still indicate the existence of private events and can be easily exploited to reveal private information such as working habits or travel routes. The artificial appearance of black boxes and large pixels also affects legitimate pattern recognition algorithms which may be used to collect traffic or usage statistics. Our plan is to develop techniques that can completely conceal the occurrence of private events without compromising legitimate data collection. For example, Figure la shows a typical surveillance video fiame with two people. If the access control database indicates that the person on the right should not be recorded, his image will be tracked and cropped as shown in Figure Ib, and completely erased from the surveillance video as shown in Figure Ie. Initial work in this direction has been performed where a stored static background is used to replace private objects [12). Nonetheless, our investigations show that static background alteration is inadequate to cope with the complex and dynamic nature of the environment. A change in illumination, shadows cast by private objects, or background motion occluded by private objects are some of the many challenges imposed by video alteration. Our efforts will initially focus on simple techniques such as improved background modeling [2] and the use of multiple cameras to eliminate shadows [6]. In the future, we plan to investigate the more-sophisticated video in-painting techniques [13] and the possibility of searching for a replacement scene from a surveillance database, inspired by our work on similarity searches in video databases [3]. While video alteration provides a powerful way to hide individuals and events, it destroys the very nature of video surveillance as a valid record of the environmental transactions. The altered video can no longer be used in a court of law to show the whereabouts of individuals. To guarantee the intrinsic legal value of video surveillance, the original version must be preserved, but in a secure way so as not to defeat the purpose of privacy protection. A. Senior et al. envisioned a system that stored multiple versions of a surveillance video, each revealing a different level of privacy information which could be retrieved only by users with the appropriate level of security clearance [9]. Storing multiple versions of the video with different levels of security, however, makes the system more susceptible to error and malicious attack. We propose a more practical and elegant solution, which is to fuse together the original private objects and the altered video using secure data hiding or watermarking. The resulting fused video is simply a standard video file, decodable by any video player and completely impervious to potential attackers. Only users with the appropriate decoder and the correct secret key can unlock the hidden information. The main technical challenge of this approach is hiding large amounts of information in a video without degrading its visual quality. Our solution involves computing a spatial perceptual mask (Figure 1d) for each frame and concealing more bits in the brighter and high textured areas. Initial results show that the watermarked video has little visual distortion (Figure Ie), at the expense ofa significant increase in file size (three to eight times