Distributed privacy for visual sensor networks via Markov shares

Visual sensor networks (VSNs) can be used to acquire visual data (i.e. images) for applications such as military reconnaissance, surveillance, and monitoring. In these applications, it is of utmost importance that visual data be protected against eavesdropping to uphold confidentiality and privacy rights. Furthermore, protection mechanisms for these sensor nodes must be efficient and robust to node capture and tampering. This paper considers a distributed approach to privacy in which highly correlated images within a dense sensor cluster are obfuscated. The particular approach, in which nodes within a cluster work together to create and transmit shares (called Markov shares) makes it necessary for an attacker to capture several correlated visual nodes and/or shares in order to gain improved semantic information of the observation area. The proposed technique does not require that the individual sensor node readings be exactly registered, nor the correlation model be known a priori. Simulation results based on a cluster of 18 nodes show: (1) most Markov shares use fewer bits per pixel than the original image hence providing compression capability; (2) a denial of service attack on a single node (e.g., corrupting a region of interest) has minimal impact on the reconstructed data at the sink; (3) five or more Markov shares need to be intercepted by an attacker before the semantic content of the desired image can be understood; and (4) authorized reconstruction of unregistered individual images with random rotation transformations up to 10 degrees is possible