Adaptive knowledge extraction and loose semantic coupling in multimedia publish/subscribe overlay networks
Zaarour, Tarek
Zaarour, Tarek
Loading...
Repository DOI
Publication Date
2022-06-14
Type
Thesis
Downloads
Citation
Abstract
The primary use of communication networks today has shifted towards content dis-tribution despite being originally designed for conversations between communicatingendpoints. The massive amounts of data generated by technologies such as the Inter-net of Things (IoT) and Social Networks have resulted in group communication beingthe prominent way to disseminate information as event notifications. The recent emer-gence of multimedia-based services and applications in the IoT has given rise to a newparadigm coined the Internet of Multimedia Things (IoMT). In the IoMT, smart het-erogeneous multimedia things interact and cooperate with one another and with otherthings connected to the Internet to facilitate multimedia based services and applications.Connecting sensing devices to the internet is at the core of many IoMT applications andplays an immediate role in facilitating the acquisition and on-the-fly processing of datafor real-time applications such as Autonomous Driving, Traffic Management, and SmartAgriculture. The scalability and interoperability offered by the publish/subscribe com-munication pattern has proved beneficial for connecting heterogeneous sensing devicesand actuators to each other and to the internet. In this thesis, I focus on offering agroup communication service that targets the dissemination of unstructured multimediacontent such as images, videos, or audio, to a diaspora of applications at internet-scale.Current implementations of distributed publish/subscribe systems take the form ofapplication-level overlays that span two main infrastructures, namely, semi-centralisedbroker overlays and fully decentralized peer-to-peer overlays. The problem at hand en-tails integrating knowledge extractors, i.e. Operators, into the event processing pipelineof distributed publish/subscribe systems. These operators are in charge of uncoveringmeaningful semantic concepts from the multimedia data being disseminated. Extracted concepts can then be evaluated against expressions submitted by subscribers as in con-ventional content-based event matching approaches. In the case of federated brokeroverlays, the closed nature of the system, and the availability of a single entity withexplicit control over nodes in the overlay allows for the integration and management ofsuch operators. We propose embedding light-weight binary filters (i.e. image classifiers)into the forwarding function of brokers organized into a federated overlay. We formu-late the shared-filer placement and ordering problem and we provision event matchingand routing algorithms for the adaptive ordering and distribution of classifier executionsalong paths towards interested subscribers.In the case of peer-to-peer overlays, the decentralized and highly scalable nature ofstructured peer-to-peer networks makes them a great fit for facilitating the interac-tion and exchange of information between dynamic and geographically dispersed au-tonomous entities. The decoupled nature of publish/subscribe systems exacerbated bythe decentralized and large-scale nature of peer-to-peer networks produces a semanticgap between publishers and subscribers. More precisely, the large semantic space ofhuman-level recognition creates a very large data space of object labels, attributes, andrelationships. The scale of the content space makes it nearly impossible for participantsto agree on a bounded set of terms for subscribers to express their exact interests. Weidentify an inherent limitation of peer-to-peer networks lying in the exact-match prop-erty of their key-based routing primitives. We propose an approximate matching modelwhere participants agree on a distributional model of word meaning that maps termsto a vector space. We overcome the exact-match limitation by proposing a novel dis-tributed lookup protocol and algorithm to construct a peer-to-peer network and routecontent. We replace conventional logical key spaces with a high-dimensional vector spacethat preserves the semantic properties of the data being mapped. We further proposemethods to partition the space, construct a semantic DHT via bootstrapping, performapproximate semantic lookup operations, and cluster nodes based on their shared inter-ests.
Funder
Publisher
NUI Galway