The constrained availability of resources on an edge analytics platform prompted the need for a trade-off between accuracy and latency by selecting suitable deep neural network (DNN) models on-the-fly. Earlier efforts either used a single powerful multi-core edge computing device or a distributed cluster of edge nodes. While the former has a high cost and power consumption, the latter incur a high communication overhead. In this paper, we propose a quality-of-service (QoS) aware video analytics platform using an edge-cluster made of low-cost devices. The edge nodes, that constitute the cluster, host heterogeneous DNN models having different configurations and number of layers. The nodes cooperate among themselves to jointly process a streaming video to achieve an optimal QoS. We formulate an optimization problem using penalty as the utility function to minimize the long-term average penalty (LTAP). We first design a DNN model recommender algorithm to minimize the LTAP and then compare it with an Oracle to show that it can achieve an LTAP with an error of 1.6 % and 9.88 % for video resolutions of 720p and 2160p respectively. We also show that the bounds on LTAP are lower and tighter for lower resolution videos compared to the higher resolution videos.

While artificial intelligence (AI) based algorithms have become an epitome of video quality analysis (VQA) for streaming services, the true power of AI is still to be harnessed for situation aware streaming in power constrained wireless networks. In this work, we propose an architecture for situation-aware streaming, that identifies important events from live-feeds of the surveillance cameras, and allocates optimal power that reduces the long-term power consumption. Real-time video surveillance is a crucial technology for smart cities, that requires deployment of large number of cameras; both road-side and aerial. The proposed architecture is designed to (i) reduce the total power consumption of all the surveillance cameras, thus reducing the greenhouse emissions, (ii) improve flight duration of aerial cameras (e.g. drones), (iii) reduce the manual searching of desired events/objects, and (iv) improve the overall QoE. It is imperative that only the important events (example, a car violating a red light) are of interest to a law enforcement officers. Hence, if the important sections of the video are received with high quality, the long-term QoE increases. The architecture has two modules, viz., a tiny neural network (having a small number of hidden layers) at the source, that incurs a small computational resource, albeit at the cost of low accuracy; and a deep neural network (DNN) (with many hidden layers) at the destination that is capable of determining events with high accuracy. We show that there exists an optimal number of frames that provides optimal QoE and is able to reduce the required power consumption of the transmitter compared to when situation awareness is not used.

Device-to-device (D2D) communication underlay cellular network improves spectral efficiency by reusing cellular resources over short low-powered links between the devices. These D2D links can be used to fetch cached contents and therefore improve the application experiences, especially for multimedia contents. In this article, we investigate the effect of caching policies on the offloading probability (also known as the hit-rate), and hence the buffering time, for video streaming applications. Designing a caching policy to maximize the hit-rate depends on several factors such as device storage capacity, instantaneous number of devices available to act as a D2D transmitter, resource allocation policy, etc. We first model the D2D links as a simple birth-death process to realize the dynamicity of the D2D links, and derive the steady-state distribution of the number of D2D links. We show that the hit-rate can be improved by increasing the caching redundancy (number of copies of a given content to be cached). Finally, through simulation we show that there exists an optimal group size and redundancy, beyond which the hit-rate do not improve.

Mobile Opportunistic Networks (MONs) are characterized by intermittent connectivity with long isolation period, and nodes following redundant transmissions for reliable message delivery. This often leads to unnecessary buffer occupancy, preventing new messages from getting replicated due to small contact duration and low bandwidth, or leading to packet drop under constrained buffer. Although attempts have been made to mitigate buffer congestion, the existing schemes are localized, are slow to react, or are specific to a routing scheme. Moreover, they rely on message exchanges to obtain buffer state/occupancy, thereby incurring additional overhead. In this paper, we first develop a generalized probabilistic forwarding model where the forwarding probability denotes the likelihood of a message to get forwarded to the encountered node. Based on the forwarding probability, we develop a congestion indicator and predict the point of congestion using the Kalman filter. Using this, a node can decide the optimal number and the exact set of messages to replicate, which leads to an optimal performance with minimal packet drop and overhead. Simulation results using a synthetic mobility model and a real-life mobility trace show that the proposed scheme outperforms the existing schemes.

Mobile opportunistic networks (MONs) are characterized by frequently changing network topology that rely on multi-copy transmission schemes to ensure the delivery of messages. However, the limited buffer capacity of mobile nodes leads to buffer overflow and head-of-line blocking. Head-of-line-blocking is caused due to short contact duration, which causes the older messages to be replicated at a faster rate than the newer messages waiting behind them. This unfair replication decreases the message delivery ratio and increases the average delivery delay. For implementation, we categorize a message as new or old and calculate the necessary maximum number of replicas required to deliver a message to the destination. If the instantaneous number of replicas in the network is less than this number, we schedule a message for replication and vice-versa. We demonstrate that our proposed scheme achieves a delivery ratio comparable to a utility-based optimal message drop algorithm with significantly lower overhead.

Device-to-device (D2D) communication helps in enhancing the capacity of the cellular network. However, the provision of video streaming in a D2D underlay network is challenging due to the dynamic and limited availability of resources especially under high mobility. Scalable video coding (SVC) allows for dynamic adjustment of video quality level according to the instantaneous network conditions, e.g., achievable data rate, player butter occupancy and user preferences.In this paper, we propose an optimal decision theory (ODT) based scheme to fill the buffers with appropriate video quality levels so as to minimize the absolute distortion under constrained stall percent. A setup with one-pair of D2D user and other cellular users (CUs) sharing the same bandwidth with different mobility patterns is used to evaluate the proposed scheme. The scheme is compared with two other schemes viz., the random allocation scheme and the greedy allocation scheme to observe the performance of the ODT based scheme.

We propose an idea to speed up instruction execution through a probabilistic approach, using the parallelism offered by quantum computers. For this, we divide the instruction set of an arbitrary quantum instruction set architecture (QISA) into separate groups and then bias certain qubits representing the group so that only the instructions within the group have a high probability of getting executed in a quantum processor. Therefore, the result generated will be the superimposition of the qubits as if all the instructions within the group were executed simultaneously. We show that we can achieve a significant design improvement compared to classical computer.

Due to the unavailability of an end-to-end path between the source and the destination, nodes in Mobile Opportunistic Networks (MON) follow a replication based strategy for message delivery. Such replications occur in bulks during intermittent and very short contact events, which often lead to buffer congestion in the relay nodes, thereby affecting the network performance. Older messages (those which have already been delivered to the destination) also tend to stay in the network for a longer time period hindering the spread of newly generated messages. Earlier works have considered buffer management techniques through local information exchange, which leads to overhead. In this work, an expression for the expected buffer occupancy under a given network scenario is obtained using the notion of a classical balls-into-bins problem. Thereafter, an estimator to estimate the buffer occupancy is designed, which can be directly used in buffer management algorithms without local information exchange. We compare the theoretical model, with estimated and simulated results, to prove the correctness of the model.

In Mobile Opportunistic Networks (MON), messages are transferred in bulk during a contact event. Such bulk transfers may cause a sudden increase in buffer occupancy, which may lead to congestion, and hence impact network performance. Apriori information of bulk transfer size distribution may help in taking proactive measures to prevent the onset of congestion. Therefore, we study the characteristic of bulk transfer size (BTS) and its impact on buffer occupancy. For this, we simulated the network using synthetic and real-life mobility traces and found that the aggregate inter-contact time (ICT) and aggregate BTS can be well approximated by the lognormal distribution. Further, we devised an estimator for mean BTS, which can be used by a receiving node to decide whether to accept or reject the bulk in order to avoid buffer overflow. We also found that an increase in expected ICT leads to an increase in expected BTS, which is confirmed using a simple linear regression model.

Security is the major issue in most of the realtime applications, besides there being a stringent requirement for such applications to provide high QoS. As security algorithms demand large computation time, the prevalent real-time packet scheduling algorithms tend to focus on meeting the deadline without having much concern for security requirements. In this paper, we propose a utilization based secured dynamic scheduling algorithm (u-SDSA) for real-time applications on grid which tries to meet both (security and scheduling) the above-mentioned requirements. Using grid of nodes or computing elements as the distributed framework for processing real-time packets, we try to ensure maximum guarantee ratio, maximum average security level, and optimal overall performance. To ensure optimal overall performance of the grid, our mechanism distributes the incoming packets for processing amongst different nodes based on utilization of the nodes. If the minimum security level of an incoming packet cannot be ensured, then it is forwarded to one of the adjacent nodes having the least utilization, provided the utilization of this nodeis less than the utilization threshold value (Uthresh). Using extensive simulation, we show that the proposed u-SDSA algorithm performs better than the existing algorithms.

Recent advances in mobile communication shows proliferation in networks formed by human carried devices known as the Pocket Switched Network (PSN). In this paper we analyze the nature of community formation since communication in such networks is highly dependent on the socializing behavior of humans. Using real world mobility traces we propose an online algorithm for the nodes to detect their community members based on the mobility parameters contact time (CT) and inter-contact time (ICT). We also derive an estimator for power-law index since CT and ICT follow power-law distribution. Based on the obtained information we propose a social based routing algorithm, named Community Aware Two-Hop routing, and compare it with generic Epidemic and Prophet routing and Bubble-Rap, a social based routing. Results show that the proposed routing protocol performs better than Bubble-Rap and achieve similar performance with respect to Epidemic and Prophet, but at a much lower cost.