We present a domain adaptation based generative framework for zero-shot learning. Our framework addresses the problem of domain shift between the seen and unseen class distributions in zero-shot learning and minimizes the shift by developing a generative model trained via adversarial domain adaptation. Our approach is based on end-to-end learning of the class distributions of seen classes and unseen classes. To enable the model to learn the class distributions of unseen classes, we parameterize these class distributions in terms of the class attribute information (which is available for both seen and unseen classes). This provides a very simple way to learn the class distribution of any unseen class, given only its class attribute information, and no labeled training data. Training this model with adversarial domain adaptation further provides robustness against the distribution mismatch between the data from seen and unseen classes. Our approach also provides a novel way for training neural net based classifiers to overcome the hubness problem in zero-shot learning. Through a comprehensive set of experiments, we show that our model yields superior accuracies as compared to various state-of-the-art zero shot learning models, on a variety of benchmark datasets. Code for the experiments is available at github.com/vkkhare/ZSL-ADA.

A concurrent prefix hijack happens when an unauthorized network originates IP prefixes of multiple other networks. Its extreme case is leaking the entire routing table, i.e., hijacking all the prefixes in the table. This is a well-known problem and there exists a preventive measure in practice to safeguard against it. However, we investigated and uncovered many concurrent prefix hijacks that didn't involve a full-table leak. We report these events and their impact on Internet routing. y correlating suspicious routing announcements and comparing it with a network's past routing announcements, we develop a method to detect a network's abnormal behavior of offending multiple other networks simultaneously. Applying the detection algorithm to BGP routing updates from 2003 through 2010, we identify five to twenty concurrent prefix hijacks every year, most of which are previously unknown to the research and operation communities at large. They typically hijack prefixes owned by a few tens of networks, last from a few minutes to a few hours, and pollute routes at most vantage points.

Content Delivery Networks (CDN) are overlay network of servers being used to deliver growing traffic demands on the Internet. As a result, CDNs are facing ever-increasing operating costs. Internet Service Providers (ISP) charge CDNs on server traffic, computed using common usage-based charging models, e.g. 95th Percentile charging. We propose Network Cost Aware Request Routing, NetReq, that assign user requests to reduce server charging volume. We compare NetReq against nearest-available server request routing in large scale simulations for both Web and multicast traffic requests. NetReq reduces charging volume for both traffic request types, thereby reducing cost. NetReq provides comparable network performance for multicast traffic by introducing end-to-end delay as a constraint in the request-routing. NetReq marginally increases network performance for Web traffic, when content maybe available at every server.

Internet Service Providers (ISPs) route traffic at the IP layer with the preference of less inter-carrier payments while Content Distribution Networks (CDNs) route traffic at the application layer with the preference of better application performance. Such mismatch of routing preferences leads to conflicts that eventually result in higher operational cost for both ISPs and CDNs. In this paper, we propose to make CDN and ISP routing mutually beneficial through ISP's non-uniform bandwidth charging and CDN's bandwidth cost-aware request routing. More specifically, ISPs charge different prices for traffic that traverses different types of inter domain links and CDNs, in routing user requests to their servers, try to minimize their ISP payments by taking the pricing information into consideration. We evaluate the solution in large scale simulations. The greedy solution presents the lowest bandwidth cost for CDNs but at the expense of network performance for users. With end-to-end delay introduced as a constraint in the optimization process, the solution maintains good network performance for users while achieving significant savings in bandwidth cost. Compared with conventional nearest-available policy in CDN request routing, our solution moves significant amount of inter domain traffic from provider routes to peer or customer routes, reducing operational costs for ISPs and CDNs.

Internet routing tables have been growing rapidly due to factors such as edge-site multihoming, traffic engineering, and disjoint address allocations. To address the routing scalability problems caused by this rapid growth, we propose an evolutionary approach that is incrementally deployable and provides immediate benefits to any adopting ASes. The basic premise of the approach is that route aggregation removes from routing tables the unnecessary topological details about remote portions of the Internet. We demonstrate that aggregation can be applied incrementally starting from local scopes within individual routers and individual ASes, and gradually expanded to the global Internet scope. The evaluation studies show that route aggregation is effective in addressing FIB scalability problems within a router and within a network.

Internet-scale dissemination of streaming contents (e.g., live sports games) can be achieved by infrastructure-based overlay multicast networks, where multicast service providers deliver the contents via dedicated servers strategically placed over the Internet. Given the huge amount of data traffic, one of the major operation costs is the ISP cost for network access. However, existing overlay multicast protocols only consider network performance metrics in building dissemination trees without taking into account the potentially high ISP cost they may incur. This paper presents a scheme, revenue-driven overlay multicast networks (ROMaN), to assign users to different servers in order to maximize the profit derived from providing multicast services. ROMaN exploits the fact that ISP charging functions are concave by assigning users to the cheapest available servers, and dynamically adjusts the assignment to accommodate the churns of group membership. The evaluation shows that ROMaN not only can reduce ISP cost substantially, but also has shorter end-to-end delay due to smaller overlay size, and the longer a user stays in the group the better the service it will receive.