Danyang Zhang

A recovery algorithm for reliable multicasting in reliable networks

Any reliable multicast protocol requires some recovery mechanism. A generic description of a recovery mechanism consists of a prioritized list of recovery servers/receivers (clients), hierarchically and/or geographically and/or randomly organized. Recovery requests are sent to the recovery clients on the list one-by-one until the recovery effort is successful. There are many recovery strategies available in literature fitting the generic description. We propose a polynomial time algorithm for choosing the recovery strategy with law recovery latency without sacrificing much bandwidth. We compared our method with two existing recovery methods, SRM (scalable reliable multicast) and RMA (reliable multicast architecture), by simulation and found that our method performs better. Although our theoretical analyses are based on a reliable network, our simulation results show that our strategy performs as well with the per link loss probability in a network up to 20% or more

Static subgroup-based source recovery for reliable multicast in reliable networks

In this paper, we consider the problem of multicasting large files over a reliable network. In a reliable network, errors are transient and any change in the multicast tree is a rare occurrence. Further, the loss of a packet is correlated in the sense that a packet lost at a link will result in a loss to all downstream recipients. We propose to partition the recipients into static subgroups during the construction of the multicast tree. Whenever a NACK is received, the source retransmits the packet to all members of the subgroup from which the NACK came. This recovery method eliminates the overhead of joining/leaving subgroups associated with dynamic recovery schemes. This scheme reduces the NACK implosion by allowing all NACK from a subgroup to be merged in one NACK. We proposed an objective function to judge the merits of static subgroup-based recovery schemes. Further, we proved that computing the optimal subgroups is NP-hard. We provide an heuristic for computing subgroups with low cost of recovery and study its performance by simulation.

Supporting time-critical clients in scalable pub-sub systems

In a pub/sub (publish/subscribe) environment like JBI (Joint Battlespace Infosphere), information/data has to be efficiently and reliably transmitted between servers and clients. However, due to various reasons such as unexpected network traffic, hardware/software failure, etc., the overlay network of the system needs to be reconfigured, thus intolerable service interruption may occur to time critical applications. This paper proposed theoretical methods for providing temporary respite to the time-critical clients while the backbone network is under repair

Energy-efficient distributed authorisation server (DAS) placement for sensor networks

Sensor networks are resource-constrained environments. Further, sensor networks are not physically secured, that is, sensor nodes may be physically captured and reverse engineered by the attacker. The goal of our research is to design algorithms for placing energy-efficient distributed authorisation servers (DASs) resilient to logical attack (hacking) and physical attack. A common method for achieving longer life for a sensor network is to deploy redundant nodes and to activate only a subset of nodes at one point of time. To the best of our knowledge, nobody addressed the problem of choosing the nodes to be activated. In this paper, we designed a set of algorithms to choose the set of nodes for forming the DAS between the command and the control centres (CCCs) and sensors collecting data. Our algorithms maximise the time required to compromise the security as well as to reduce the energy requirement for both authorisation and communication.

Automatic detailed localization of facial features

We propose a complete framework for automatic detailed facial feature localization. Feature points and contours of the eyes, the nose, the mouth and the chin are of interest. Face detection is performed followed by the region detection that locates a rough bounding box of each facial component, and detailed features are then extracted within each bounding box. Since the feature points lie on the shape contours, we start from shape contour extraction, and then detect the feature points from the extracted contours. Experimental results show the robustness and accuracy of our methods. The main application of our work is automatic diagnosis based on facial features.

A generic content-based image search model and its key factors

Despite the popularity of text-based image search engines like Google Image Search and Yahoo Image Search, their incapability of catching the real content/ sense of images makes content-based image search engines necessary. However, currently efficient and popular content-based image search engines are rare. This may be due to unfriendly user interface, low precision ratio, low recall ratio, etc. This paper presents a generic content-based image search model and briefly analyzes its key factors.

A server relay recovery strategy (SRRS) for reliable multicast over Internet-like topologies

For the last decade, a number of hierarchical reliable multicast schemes have been proposed in the literature due to their efficiency and scalability. They normally partition the group members into local groups and allocate one server for each local group to detect and recover packet loss. One common problem is how the server recovers the lost packet in the case that the server itself does not receive it. The usual solution is either to send requests to some upstream servers/source or to construct a tree-based hierarchy to organize the servers for recovery. However, these solutions may lead to long recovery latency. This paper presents a server relay recovery strategy (SRRS) for servers to achieve low recovery latency and low bandwidth usage for recovery. The basic idea is for each server to distributively compute and maintain a prioritized peer list that leads to minimal expected relay recovery latency. Our simulation results show that when the server distribution is not dense, SRRS outperforms SRM (scalable reliable multicast) and RMA (reliable multicast architecture) on both recovery latency and bandwidth usage in Internet-like topologies.