Codebook design for ASR systems using custom arithmetic units

This paper studies the design, implementation, and evaluation of a new universal adversarial attack targeting automatic speech recognition systems in a black-box setting. A genetic algorithm optimizes universal perturbations consisting of short noise bursts that cause mistranscriptions by balancing text similarity (character error rate) and perceptual audio similarity (Mel energy distance) to keep the noise minimally intrusive. Experiments are conducted on the models Wav2Vec 2.0 and OpenAI’s Whisper to investigate the attack’s performance under varying parameters such as noise volumes, number of audio files in the training set, and for the standard English Librispeech dataset, as well as a synthetic maritime dataset that contains more homogeneous data. We expose vulnerabilities in state-of-the-art ASR systems and the risks of attacks on safety-critical applications, such as maritime radio communication. We demonstrate that our attack is highly successful, and even an attack trained on a single input works universally. Whisper proves to be more robust against these attacks. We find that universal perturbations generalize better when trained on data more similar to the test set. A semantic defense is developed that presents a novel way to detect the attack. To our knowledge, our work represents the first universal black-box attack against ASR models.

Ahstract— With the rapid development of vehicular applications and mobile devices, demands for resources to process time-critical and computation-intensive tasks are increasingly prominent. In this paper, we propose a two-layer Vehicular Fog Computing (VFC) architecture, including the client layer and the fog layer. Vehicles may generate tasks as clients, which are further assigned to the nodes in the fog layer for processing. The fog layer aggregates available resources of vehicles and infrastructures by exploiting their communication, computation and storage capabilities. Each task requires certain amount of resources for processing at the fog nodes. We formulate a distributed task allocation (DTA) problem, which takes deadline, vehicle mobility and fog capacity into consideration, and aims at maximizing the overall resource utilization of system, via the cooperation of vehicles and fog nodes. We linearize DTA into a 0–1 integer linear programming (ILP) problem to obtain the optimal solution. Further, we design a heuristic algorithm to obtain near-optimal performance with low computational overhead, which decomposes DTA into two subprocess and schedules tasks in each fog node independently. Finally, we build the simulation model and conduct a series of experiments based on real-world vehicle trajectories, which demonstrate the effectiveness and scalability of the proposed algorithm.

A Novel Lyapunov based Dynamic Resource Allocation for UAVs-assisted Edge Computing

In this paper we first show that the Wi-Fi Protected Setup (WPS) protocol (used by Wi-Fi Direct, the de facto adhoc communication mechanism for smartphones and mobile devices) is vulnerable to a brute-force or dictionary attack. To defend against these attacks, we propose the idea of using contextual information (i.e., data obtained from mobile device's sensors) to establish a long (128 bits) secure session key between two Wi-Fi Direct enabled devices, instead of using the keypad. Our solution, Session Key Generated from Sensors (SekGens) employs three phases. In the Quantization Phase, the key is iteratively generated based on different sensors' data. In the Reconciliation Phase, the two devices eliminate minor differences in the bits of their keys by using the Cascade reconciliation mechanism. In the Privacy-Amplification-and-Hashing Phase, the two devices omit all bits exposed during the reconciliation phase and apply hashing to the remaining secret bits. SekGens is implemented and evaluated by modifying the Android kernel code responsible for WPS in Google Nexus 5 and Samsung Galaxy S2 smartphones. The results show that SekGens generates keys with low mismatch ratio (less than 3%), at a fast rate (~20 bits/sec), and with high entropy (~92%).

Fog computing is an energy-efficient and cost-effective paradigm to help alleviate the pressure of resource-constrained mobile devices (MDs) running computation-intensive applications. In this article, we investigate the joint task partitioning and power control problem in a fog computing network with multiple MDs and fog devices (FDs), where each MD has to complete a periodic computation task under the constraints of delay and energy consumption. Each task can be partitioned into multiple subtasks and offloaded to the FDs according to the task partition strategy and transmission power strategy to reduce task execution delay and energy consumption. To this end, we present a multiagent deep deterministic policy gradient (MADDPG)-based task offloading algorithm for MDs to maximize the long-term system utility including the execution delay and energy consumption. Each MD inputs the local information, e.g., the task requirements, the available communication, and computation resources of the FDs, the computation resources, and the battery level of the MD into a distributed actor network to generate a task offloading policy, while a centralized critic network is used to update the weights of the actor networks to improve offloading performance. Numerical simulation results demonstrate the effectiveness of the proposed scheme in improving the system utility, reducing the average execution delay as well as the average energy consumption.

The paper introduced a speech feature grouping algorithm for the speech recognition system based traditional Markov in accordance with the large computation of the traditional hidden Markov model and the Viterbi algorithm as well as the Gaussian mixture distribution probability. For the speech characteristic parameters, clustering was executed by K-Means algorithm on the basis of the first and second segmentation, and then obtained the grouped characteristic parameters and the parameters to be grouped, and the speech samples can be divided into different characteristic group according to these two parameters. On this basis, a grouping training algorithm was proposed by using the redundant, which improved the accuracy of grouping the speech characteristic by clustering algorithm. Compared with the traditional HMM method, the amount of calculation can be reduced more than 60% in the case of ensuring the speech recognition rate. In the traditional HMM speech recognition system, and for the isolated word, the classic Viterbi algorithm perfectly solved the decoding problem with iterative approach in the mathematics. But in the speech recognition systems, for the large vocabulary in the practical application of acoustic decoding, the amount of computation that the algorithm required is considerable. Assuming the vocabulary that a speech recognition system can recognize is 1000, then to establish a model for each word, and assuming they have the same number of states N, in order to facilitate the estimation and calculation, these models must be connected to a large model. So, the state number of the large model is 1000 times larger than that of original model. Because the orders of computation that the Viterbi algorithm required is N2T (T is the number of input speech frame), the computation amount of Viterbi algorithm of large models increased four orders than that of the original phoneme model. Moreover, according to the experiment, it was found that the computation was mainly used for calculating the multi- dimensional (mixture) Gaussian distribution probability of the obtained observation vector for each state of every frame. Therefore, the real-time requirement for the small mobile device with large vocabulary HMM speech recognition system is difficult to meet because of the limit computing capability. Contrary to the large amount of computation for the Viterbi algorithm and the Gaussian mixture distribution, the paper presented a grouping model of speech characteristics, before the matching calculation of the input speech done, the grouping judgment was executed firstly and obtained the group that the input speech belonged to, then to calculate the HMM parameters of input speech matching with that of all the speech in the same group. Then combined with the improved HMM speech recognition system, the efficiency can be improved.

Real-time computing provides insightful ways to explore the optimization in resource usages, especially from the time point of view. Nevertheless, real-time task scheduling is recognized by its high complexity when there are non-preemptive shared resources and multiple processors. When more and more practical factors in system designs are considered, such as energy consumption and memory allocation, even some sub-problems in real-time task scheduling become intractable. Although people often criticize various artificial assumptions in real-time task scheduling, they have to admit that ideas in real-time computing and their extensions, such as tradeoff in cost, performance, energy, and even the quality of service, can be applied to multi-dimensional optimization in system designs. In this direction, we witness the rapid development of the embedded system industry and join the task force in system designs, especially mobile devices and non-volatile memory systems. Resource management on mobile devices, with a special emphasis on user experience, should not only consider the response time but also the visual perception of users. Non-volatile memory has also blurred the boundary between the memory and the storage. It enables certain unified considerations of the main memory and storage and also in-memory computing. It shows the ways to break the boundaries between hardware and software layers and have better integration of computing and memory/storage units. The advances in mobile systems and memory innovations inspire the evolution of embedded system designs and have also brought us insights to solutions regarding how systems should be restructured and how computing should be done. They might also provide their feedback to real-time computing and even shape the future direction of real-time computing in various innovative ways.

In distributed speech recognition, the speech features are computed on a mobile device, compressed, and sent over a network to a speech recognition server, where the Viterbi search and hidden Markov modeling take place. We examine some error concealment methods for distributed speech recognition over burst error channels. We consider interpolation and interleaving, and we present a novel use of the stochastic weighted Viterbi recognition algorithm to increase robustness against interpolated features. We examine interleaving at both the frame level and codebook index level. Channel errors are simulated using a Gilbert model, and the performance of our algorithm is compared with other techniques, including the ETSI DSR standard, on a digits task and a large vocabulary task. Coupled with interleaving and interpolation, our algorithm can provide accuracy as high as 96.7% on a digit recognition task during an average bit error probability of 1/20. On the more difficult WSJ task, the accuracy without bit errors is 85.7%. Using our algorithm, we can achieve 82.9% accuracy with an average bit error probability of 1/30.

Task scheduling is one of the very crucial facets of cloud computing. The task scheduling method must assign jobs to virtual machines. In cloud computing, task scheduling includes a frontal influence on a system's resource utilization and operational costs. Diverse meta-heuristic algorithms, in addition to their modifications, have been developed to improve the efficiency of task executions in the cloud. In this paper, a multiobjective optimization model is applied using the metaheuristics cuckoo search optimization algorithm (MCSO) to enhance the performance of a cloud system with limited computing resources while minimizing the time and cost. Finally, we analyze the performance of the proposed MCSO with the existing methods, such as Bee Life Algorithm (BLA), A Time–Cost aware Scheduling (TCaS) algorithm, Modified Particle Swarm Optimization (MPSO), and Round Robin (RR), for the evaluation metrics makespan and cost. Based on the outcomes of the experiments, it can be inferred that the proposed MCSO provides essential schedule jobs with the shortest makespan and average cost. Index Terms— Task Scheduling, Multiobjective, Cloud Computing, Cost, Makespan.

Path computation in a network is dependent on the network's processes and resource usage pattern. While distributed traffic control methods improve the scalability of a system, their topology and link state conditions may influence the sub-optimal path computation. Herein, we present Pathfinder, an application-aware distributed path computation model. The proposed model framework can improve path computation functions through software-defined network controls. In the paper, we first analyse the key issues in distributed path computation functions and then present Pathfinder's system architecture, followed by its design principles and orchestration environment. Furthermore, we evaluate our system's performance by comparing it with FreeFlow and Prune-Dijk techniques. Our results demonstrate that Pathfinder outperforms these two techniques and delivers significant improvement in the system's resource utilisation behaviour.

Fast multi-resource allocation with patterns in large scale cloud data center

Fair job-scheduling problem in virtualized cloud computing environments has been a focus of research recently, yet this has not received adequate attention in the context of private cloud, where each physical machine hosts a set of virtual machines, while each of virtual machine supports several tasks running on it. In this paper, we introduce a rigorous definition of dominant resource share of tasks and virtual machines, and design a two stage dominant resource fairness scheme that investigate the problem of optimally utilizing a limited amount of a resource (e.g., CPU, memory, bandwidth, or storage) while maximizing application performance and present the inherent tradeoff between resource utilization and fairness in the job-scheduling process. We design a simple heuristic that implements our model in real-world system. Large-scale simulations driven by traces of our cloud platform show that the proposed job scheduling mechanism significantly improves the cloud system's resource utilization, yet with a substantial reduction of job completion times.

In mobile communication system, the traffic is time-varying and unpredictable, so there will be some hot-spot cells. It means that, the number of service request is much higher than an acceptable level in some cells, which may result in call-blocking with a large probability. But in other cells of system there are still resources available to serve more users. How to avoid the imbalance to effectively improve the system's resource utilization is what load balancing need to solve. In this paper, a new kind of event-trigger mechanism using cell load status information exchange between eNBs via X2 interface is presented. Additionally, based on it, a novel load balancing algorithm, named TSLB(two-staged load balancing), for LTE systems is put forward. Using this algorithm, eNBs can start corresponding load balancing procedure accurately according to the cell load status in real-time without any manual intervention. Simulation results show that the method is valid and it can increase the system throughput significantly as well as decrease the network delay.

This paper reviews the design and implementation of intelligent tutoring systems (ITS) on mobile devices, with a special focus on the technical challenges and students' experience. A Sudoku ITS is used for illustration. Unquestionably, the small display and limited processing power and storage of mobile devices pose the challenges on algorithm design and user interface design. To meet the challenges, the Sudoku ITS adopts a thin client with a neat user interface connecting to an external database. Customised and immediate instructions are provided to help students acquire skills and develop strategies in puzzle solving. Positive feedbacks from students are received.

A new nano-based architectural design of multiple-stream convolutional homeomorphic error-control coding will be conducted, and a corresponding hierarchical implementation of important class of the homeomorphic Viterbi algorithm within convolutional homeomorphic error-control coding using lattice networks via nano carbon-based field emission controlled-switching will be achieved. Error-correction coding is highly important in modern wireless networking and digital data transaction since channel coding is required to counteract data errors that are encountered in wireless data networking due to the corresponding existence of unavoidable channel noise. Further, the new lattice nano-based implementation will be useful for enhancing the error-control system performance such as the corresponding enhancements within error-correcting capability, regular synthesis, speed improvement and the minimization of power consumption. Logic homeomorphism describes properties-preserving logic mapping that is intrinsically bijective. Homeomorphic property in error-control coding is important since it is shown that the homeomorphism relationship between multiple-streams of data can be used for further correction of errors that are uncorrectable using the implemented decoding algorithm such as in the case of triple-errors that are uncorrectable using the classical Viterbi algorithm. Applications of the new regular nano-based homeomorphic architecture include low-power design of circuits and systems for enhanced and more reliable wireless data networking and transaction.