The rapid growth of vehicle networks and infrastructure has increased the need for intelligent accident risk prediction and real-time safety information sharing in the Internet of Vehicles (IoV). Accurately predicting accidents is crucial for improving road safety in increasingly congested and complex traffic scenarios. This paper proposes a novel Deep Learning (DL) architecture, the Traffic Safety Graph Attention Network (TSGAN), which advances the state of the art in predicting accident risk and issuing real-time alerts. TSGAN integrates a Spatial Graph Attention Module (SGAM) for dynamic spatial modelling and a Temporal Convolution Module (TCM) for multi-scale temporal pattern extraction. A key innovation is the Dual-Stage Risk Prediction model, which estimates both accident probability and severity, enabling more nuanced risk assessment. Additionally, the framework includes an Impact-Aware Risk Dissemination Module that generates context-aware alerts and optimized through a Fog-based Deep Reinforcement Learning (FOG-DRL) model enhanced with a Quantum Neural Network (QNN) driven policy. Evaluations on real-world IoV data and the NS-3 simulator demonstrate that the TSGAN outperforms baseline methods in prediction accuracy, alert timeliness, and network efficiency.

High volumes of data are being generated by various scientific research organizations, along with growing technologies like cloud computing, the Internet of Things (IoT), and large-scale data analytics. This results in a growing need for efficient mechanisms to transfer data across geographically distributed locations. Object-based big data transfer systems are frequently used to manage these massive volumes of data due to their scalability and adaptability. However, effective utilization of network resources while minimizing latency remains a significant challenge, particularly in heterogeneous environments where network conditions and data characteristics vary widely. Data deduplication, a well-known technique used in storage systems to reduce the amount of data being stored, can be leveraged to optimize data transfer efficiency by eliminating the transfer of redundant blocks of data. However, traditional deduplication techniques often fail to dynamically adapt to the varying conditions of big data environments, resulting in suboptimal performance. In this paper, we propose a latency-aware deduplication technique specifically designed for object-based big data transfer systems. By adapting deduplication processes in real-time based on contextual factors such as network bandwidth and data characteristics, the proposed framework aims to reduce data transfer latency. The experimental results demonstrated that the proposed framework significantly improves the data transfer efficiency, especially under high redundancy and low network bandwidth conditions. At 80 % redundancy level, the proposed framework exhibits over 80 % reduction in end-to-end data transfer time compared to the baseline approach, which does not employ deduplication. However, due to the dynamic nature of the deduplication aggressiveness, the proposed framework demands higher CPU resources. This overhead is compensated by significant improvement in end-to-end data transfers, thereby confirming the practicality of the proposed framework for largescale and latency-sensitive data transfers.

Preserving semantic equivalence between source text and its translation remains a formidable challenge in machine translation. Despite recent advancements, neural machine translation (NMT) models often fall short due to their over-reliance on word-level alignment, a limitation exacerbated by the use of cross-entropy loss. This paper introduces Rosetta, a novel framework that revisits the semantic awareness of NMT Systems. Our approach builds upon the Transformer architecture, incorporating stochastic layer selection, multi-branch attention, and group fusion mechanisms to facilitate a more nuanced understanding of context and meaning. Central to our framework is the utilization of multilingual pre-trained language models (PLMs) to reinforce semantic consistency between the source text and its translation in the semantic space. We evaluate our approach on the IWSLT’14 translation dataset. Our model achieves state-of-the-art performance on the German-English translation task, surpassing existing models with a BLEU score of 39.12. To assess the semantic equivalence of translations, we employ GPT-4 as an independent evaluator. The results demonstrate that our approach surpasses literal translation, excelling in preserving semantic similarity.

The task of Sign Language Production (SLP) in machine learning involves converting text-based spoken language into corresponding sign language expressions. Sign language conveys meaning through the continuous movement of multiple articulators, including manual and non-manual channels. However, most current Transformer-based SLP models convert these multi-channel sign poses into a unified feature representation, ignoring the inherent structural correlations between channels. This paper introduces a novel approach called MCST-Transformer for skeletal sign language production. It employs multi-channel spatial attention to capture correlations across various channels within each frame, and temporal attention to learn sequential dependencies for each channel over time. Additionally, the paper explores and experiments with multiple fusion techniques to combine the spatial and temporal representations into naturalistic sign sequences. To validate the effectiveness of the proposed MCST-Transformer model and its constituent components, extensive experiments were conducted on two benchmark sign language datasets from diverse cultures. The results demonstrate that this new approach outperforms state-of-the-art models on both datasets.

Abstract Sign Language Production (SLP) refers to the task of translating textural forms of spoken language into corresponding sign language expressions. Sign languages convey meaning by means of multiple asynchronous articulators, including manual and non‐manual information channels. Recent deep learning‐based SLP models directly generate the full‐articulatory sign sequence from the text input in an end‐to‐end manner. However, these models largely down weight the importance of subtle differences in the manual articulation due to the effect of regression to the mean. To explore these neglected aspects, an efficient cascade dual‐decoder Transformer (CasDual‐Transformer) for SLP is proposed to learn, successively, two mappings SLP hand : Text → Hand pose and SLP sign : Text → Sign pose , utilising an attention‐based alignment module that fuses the hand and sign features from previous time steps to predict more expressive sign pose at the current time step. In addition, to provide more efficacious guidance, a novel spatio‐temporal loss to penalise shape dissimilarity and temporal distortions of produced sequences is introduced. Experimental studies are performed on two benchmark sign language datasets from distinct cultures to verify the performance of the proposed model. Both quantitative and qualitative results show that the authors’ model demonstrates competitive performance compared to state‐of‐the‐art models, and in some cases, achieves considerable improvements over them.

Large scale machine learning (ML) and deep learning (DL) platforms face challenges when integrated with deduplication enabled storage clusters. In the quest to achieve smart and efficient storage utilization, removal of duplicate data introduces bottlenecks, since deduplication alters the I/O transaction layout of the storage system. Therefore, it is critical to address such deduplication overhead for acceleration of ML/DL computation in deduplication storage. Existing state of the art ML/DL storage solutions such as Alluxio and AutoCache adopt non deduplication-aware caching mechanisms, which lacks the much needed performance boost when adopted in deduplication enabled ML/DL clusters. In this paper, we introduceRedup, which eliminates the performance drop caused by enabling deduplication in ML/DL storage clusters. At the core, is aRedupCaching Manager (RDCM), composed of a 2-tier deduplication layout-aware caching mechanism. The RDCM provides an abstraction of the underlying deduplication storage layout to ML/DL applications and provisions a decoupled acceleration of object reconstruction during ML/DL read operations. OurRedupevaluation shows negligible performance drop in ML/DL training performances as compared to a cluster without deduplication, whilst significantly outperforming Alluxio and AutoCache in terms of various performance metrics.

Abstract Dense activity detection is a subtask of activity detection that aims to localise and identify multiple human activities in video clips. Existing methods adopt offline frameworks that require video frames to be available when activity detection begins. These offline methods are unable to be applied to online scenarios. An online framework is proposed for dense activity detection. The framework has two stages: warm‐up and detection. Warm‐up is the initialisation of dense activity detection, which generates a contextual model called an online aggregated‐event. After that, the method moves into the detection stage, which consists of two modules: coarse label prediction and refined label prediction. Coarse label prediction predicts activity labels by taking the online aggregated‐event as a priori; then, prediction is refined by two techniques, human–object interaction detection and online relation reasoning. The proposed method is evaluated using two dense activity datasets: Charades and AVA. The experimental results show that the proposed method has better performance than existing offline methods after the whole video input is added to the algorithm.

All-flash array is a popular memory device available for use in modern high-performance storage systems. Compared with other types of devices such as DRAM, NVRAM, and EEPROM, flash array combines the best features: shock resistance, low cost, low power consumption, and fast access. Moreover, the ever-increasing density of flash memory has led to a dramatic increase in the capacity, which allows the storage of large volume of data. However, flash memory is not optimal for managing a large number of small files because: 1. the small and random write operation is inefficient in flash memory; 2. massive metadata information occupies a significant portion of the namespace, which is relatively limited or scarce in big data storage systems. This paper introduces a novel approach, hash partitioning-based file compaction (HFC), to improve the efficiency of storing and accessing small files in all-flash storage systems. HFC consists of a file compaction tool and an access interface. The compaction tool merges a group (usually a directory) of small files into a set of "big files" to reduce the metadata required to be maintained in the on-chip memory. The data locality and tree structure of those small files are preserved. The access interface is designed to provide transparent access to the small files in the HFC big files. Experimental results confirm that the proposed method significantly enhances the efficiency of managing massive amounts of small files in flash memory in terms of namespace usage and access speed.

The task of referring expression comprehension (REC) is to localise an image region of a specific object described by a natural language expression, and all existing REC methods assume that the object described by the referring expression must be located in the given image. However, this assumption is not correct in some real applications. For example, a visually impaired user might tell his robot ‘please take the laptop on the table to me’. In fact, the laptop is not on the table anymore. To address this problem, the authors propose a novel REC model to deal with the situation where expression‐image mismatching occurs and explain the mismatching by linguistic feedback. The authors' REC model consists of four modules: the expression parsing module, the entity detection module, the relationship detection module, and the matching detection module. They built a data set called NP‐RefCOCO+ from RefCOCO+ including both positive samples and negative samples. The positive samples are original expression‐image pairs in RefCOCO+. The negative samples are the expression‐image pairs in RefCOCO+, whose expressions are replaced. They evaluate the model on NP‐RefCOCO+ and the experimental results show the advantages of their method for dealing with the problem of expression‐image mismatching.

Apache Spark (Spark) is a unified analytics engine for large-scale data processing. Unlike traditional data processing engines like Hadoop, Spark is a framework that caches data in memory. Therefore, memory management in Spark is importance. However, there are several factors that interfere with memory management. First, if users want to cache data in memory, they need to choose their own storage level. In this case, if they do not select the optimal storage level, Spark will be put a heavy burden on memory. Next, users need to select the ratio for spark memory directly within Spark. If they do not choose optimal ratio for spark memory, garbage collection overheads will be incurred. In this poster, we propose DSMM that dynamically select the above factors on the system for memory management. Our experimental result shows 13% execution time improvement as compared to standard Spark.

A lot of previous works on Map-Reduce improved job completion performance through implementing additional instrumentation modules which collects system level information for making scheduling decisions. However the extra instrumentation may not scale well with increasing number of task-trackers. To this end, we design PADS, a lightweight scheduler which uses time prediction to schedule tasks without additional instrumentation modules. Results shows PADS improves performance by 6%, 12%, and 9% as compared to ESAMR, LA, and DDAS respectively.

Even though flash memory Solid State Drives (FSSDs) outperformed traditional Hard Disk Drives (HDDs), they are still failing to reduce performance gap between microprocessors and storage in computer systems regardless of available high bandwidth. To alleviate this, we propose implementing PCM as main memory in SSDs to replace flash memory. In particular, we present a PCM File Translation Layer (PhaseFTL) that can efficiently manage address translations from host file system to PCM while allowing PCM memory blocks to wear down evenly. PhaseFTL hides PCM's constrains and does not suffer from cache miss because it's address translations are directly linked to the entire mapping table stored on fast PCM main memory.

Generative adversarial networks (GANs) are well known both for being unstable to train and for the problem of mode collapse, particularly when trained on data collections containing a diverse set of visual objects. This study introduces an adaptive hyper-parameter learning procedure for GANs as an alternative to the existing static approach. The proposed procedure is designed to mitigate the impact of instability and saturation in the original by dynamically adjusting the ratio of the training steps of both the generator and discriminator. To accomplish this, we track and analyze stable training curves of relatively narrow datasets and use them as the target fitting lines when training more diverse data collections. Experimental results show that the proposed model improves the stability and generates more realistic images.

A top‐ k spatial keyword (T k Sk) query ranks objects based on the distance to the query location and textual relevance to the query keywords. Several solutions have been proposed for top‐ k spatial keyword queries. However, most of the studies focus on Euclidean space or only investigate the snapshot queries where both the query and data object are static. A few algorithms study T k Sk queries in undirected road networks where each edge is undirected and the distance between two points is the length of the shortest path connecting them. However, T k Sk queries have not been thoroughly investigated in directed and dynamic spatial networks where each edge has a particular orientation and its weight changes according to the traffic conditions. Therefore, in this study, we address this problem by presenting a new method, called COSK, for processing continuous top‐k spatial keyword queries for moving queries in directed and dynamic road networks. We first propose an efficient framework to process snapshot T k SK queries. Furthermore, we propose a safe‐exit‐based approach to monitor the validity of the results for moving T k SK queries. Our experimental results demonstrate that COSK significantly outperforms existing techniques in terms of query processing time and communication cost.

Document novelty detection is a concept learning problem wherein the system gains its knowledge only from the positive documents under a concept and with that limited knowledge it attempts to detect the negative cases. This work focuses on learning author style as a concept from the given set of do cuments, particularly emails. Since author attribution for shorter texts such as emails is more complex compared to larger documents, the techniques originally used for the large documents prove inefficient for short texts. To address this shortcoming of existing algorithms in detecting aberration in author style, we have proposed a graph-model based technique for feature set extraction from short documents. Given the extracted feature set, we have also developed two probability based text representation schemes that could best represent a text document to an underlying one-class SVM classifier. The proposed models have been compared and evaluated on the public Enron email dataset. Applying graph based feature set extraction technique in combination with the inclusive compound probability based text representation has proved to be very efficient. The generality of the proposed method allows the approach to be applicable to all kind of text documents including emails.

The development of wireless sensor networks can be considered as the beginning of a new generation of applications. Authenticity of communicating entities is essential for the success of wireless sensor networks. Authentication in wireless sensor networks is always a challenging task due to broadcast nature of the transmission medium. Sensor nodes are usually resource constrained with respect to energy, memory, and computation and communication capabilities. It is not possible for each node to authenticate all incoming request messages, whether these request messages are from authorized or unauthorized nodes. Any malicious node can flood the network by sending messages repeatedly for creating denial of service attack, which will eventually bring down the whole network. In this paper, a lightweight authentication scheme named as Biphase Authentication Scheme (BAS) is presented for wireless sensor networks. This scheme provides initial small scale authentication for the request messages entering wireless sensor networks and resistance against denial of service attacks.

In this paper, we focus on read operations in flash memory, which have received less attention than write operations. To reduce read latency, we propose a read-aware dynamic allocation mechanism for multichannel solid-state devices. The proposed mechanism enables read operations to be executed immediately by reserving the resources of channels, packages, and dies dedicated to read operations. This is done in collaboration with an internal device queue, in which write operations are freely routed to their proper addresses while maintaining the merits of a dynamic allocation mechanism. Our read-aware mechanism reduces read latency by avoiding read operation conflicts when trying to access resources already occupied by preissued write operations. The experimental results show that, with our proposed mechanism, read latency decreases by up to 32.5% with an increase of write latency of up to 6.8% in real traces while providing high compatibility with existing dynamic allocation schemes.

Big data storage and sharing are becoming the major demand of the community. To overcome such issues, virtually unified data facilities are being presented with geodistributed data centers by providing the user with the single unified namespace. These unified data storage facilities lack efficient storage and analysis of data. To address these shortcomings in such unified data facilities, we designed and implemented time optimization model which minimizes the job execution time whereby selecting optimal data center with constraints of storage, computational and network bandwidth among all data centers. Our extensive simulation results show that our model provides the optimal decision that leads to minimal end-to-end data placement and analysis times.

Flash Translation Layer (FTL) performs virtual-to-physical address translations and hides the erase-before-write characteristics of Flash. Pure page mapped FTL, which maintains page-level address mappings, is known as the most efficient FTL. However, its huge SRAM requirement to load the entire mapping table limited adoption of its use. In order to reduce SRAM space utilization while maintaining comparable performance, we can selectively cache page-level address mappings into a small SRAM. However, the performance of this approach is limited by miss ratio of cached mapping table (CMT) on SRAM. In this paper, we propose a replica approach of the page-mapped FTL on flash, called Replica to minimize the performance penalty of CMT miss.

Electronic mails have increasingly replaced all written modes of communications for important correspondences including personal and business transactions. An e-mail is given equal significance as a signed document. Hence email impersonation through compromised accounts has become a major threat. In this paper, we have proposed an email style acquisition and classification model for authorship attribution that serves as an effective tool to prevent and detect email impersonation. The proposed model gains knowledge of the author's email style by being trained only with the sample email texts of the author and then identifies if a given email text is a legitimate email of the author or not. Extracting the significant features that represent an author's style from the available concise emails is a big challenge in email authorship attribution. We have proposed to use a graph-based model to precisely extract the unique feature set of the author. We have used one-class SVM classifier to deal with the single-class sample data that consists of only true positive samples. Two classification models have been designed and compared. The first one is a probability model which is based on the probability of occurrence of a feature in the specific email. The second technique is based on inclusive compound probability of a feature to appear in a sentence of an email. Both the models have been evaluated against the public Enron dataset.

The use of MLC/TLC (Multiple/Triple Level Cell) flash memory increases bit error rate, and declines its reliability. To remedy this loss, the Redundancy Array of Inexpensive Disk (RAID) have been widely used to enhance the reliability of the Hard Disk Drive (HDD) and the Solid State Drive (SSD). RAID 5 and RAID 6 ensure high reliability among the various RAID techniques. These RAID techniques exploit parity to recover failures, it is updated whenever data renewed. The RAID 5 technique contains separated parity in a page of different stripes. In the RAID 6 technique, however, parity is written to double. So, RAID 6 guarantees more reliability. These RAID techniques enhance reliability, stability and data recovery capability in SSD. In this paper, we propose the dual RAID technique to use both RAID 5 and RAID 6 in a particular way depending on the data reliability. Reliability of data is divided to relatively high and low, these allows to be determined by user. At this time, data which requires high reliability is managed by the RAID 6 technique, and data which requires low reliability is managed by the RAID 5 technique. The purpose of this technique is to improve data recovery capability and I/O performance in SSD. This technique is evaluated by the trace-driven simulator with Financial1, Financial2, Exchange, and MSN traces. We confirm that the dual RAID technique improves I/O performance with ensuring high reliability.

In this paper, we studied the problem of reverse k nearest neighbors (RkNN) in directed road network, where a road segment can have a particular orientation. A RNN query returns a set of data objects that take query point as their nearest neighbor. Although, much research has been done for RNN in Euclidean and undirected network space, very less attention has been paid to directed road network, where network distances are not symmetric. In this paper, we provided pruning rules which are used to minimize the network expansion while searching for the result of a RNN query. Based on these pruning rules we provide an algorithm named SWIFT for answering RNN queries in static directed road network. We evaluated SWIFT on a real world road network and our experimental results show that SWIFT significantly outperforms the naïve algorithm in terms of computational cost.

Flash memory offers attractive features, such as non-volatile, shock resistance, fast access and low power consumption for data storage. However, it has one main drawback of requiring an erase before updating the contents. Furthermore, the flash memory can only be erased for a limited number of times. These characteristics are controlled by a software layer called the flash translation layer (FTL). FTL efficiently manages read, write, and erase operations to enhance the overall performance, and considers wear-leveling to prolong the durability of flash memory. In this article, we identify the logical sector numbers corresponding to random data, termed as hot-LSNs, and distribute them to all available blocks without degrading the performance of the flash memory. From our evaluation, we found that the extra erase operations for distributing the hot-LSNs are very low compared to the overall performance. Even though Hot-LSNs Distributing Wear-Leveling Algorithm (Hot-DL) incorporates wear-leveling in the performance enhancing algorithm, Hot-DL only requires approximately 0.015% of extra erase operations compared to previous well-optimized performance enhancing algorithms, shared buffer scheme.

Flash memory is used widely in the data storage market, particularly low-price MultiLevel Cell (MLC) flash memory, which has been adopted by large-scale storage systems despite its low performance. To overcome the poor performance of MLC flash memory, a system architecture has been designed to optimize chip-level parallelism. This design increases the size of the page unit and the block unit, thereby simultaneously executing operations on multiple chips. Unfortunately, its Flash Translation Layer (FTL) generates many unused sectors in each page, which leads to unnecessary write operations. Furthermore, it reuses an earlier log block scheme, although it generates many erase operations because of its low space utilization. To solve these problems, we propose a hybrid associative FTL (Hybrid-FTL) to enhance the performance of the chip-level parallel flash memory system. Hybrid-FTL reduces the number of write operations by utilizing all of the unused sectors. Furthermore, it reduces the overall number of erase operations by classifying data as hot, cold, or fragment data. Hybrid-FTL requires less mapping information in the DRAM and in the flash memory compared with previous FTL algorithms.

FTL (Flash Translation Layer) hides details of flash memory, providing file systems with an abstract view of the flash memory. For NAND flash memory, some previous researches have achieved dramatic performance enhancement by adopting log-based FTL, which records time-consuming write operations in log blocks, rather than executes them immediately. Log block replacement scheme plays an essential role in this method, due to the limitation of pre-reserved log block space, this method entails selecting some victims from the existing blocks and re-use them for newly issued operations. However, simple replacement algorithms are vulnerable to select such log blocks that will be used soon, which causes performance degradation. In this paper we propose a smarter log block replacement scheme to alleviate this problem by keeping busy log blocks from being selected, based on profiling and analyzing log block status. We show that our scheme reduces unnecessary time-consuming write operations and achieves performance improvement especially for the applications having intensive locality.

Flash memory is being rapidly deployed as data storage for mobile devices such as PDAs, MP3 players, mobile phones, and digital cameras, mainly because of its low electronic power, nonvolatile storage, high performance, physical stability, and portability. One disadvantage of flash memory is that prewritten data cannot be dynamically overwritten. Before overwriting prewritten data, a time-consuming erase operation on the used blocks must precede, which significantly degrades the overall write performance of flash memory. In order to solve this “erase-before-write” problem, the flash memory controller can be integrated with a software module, called “flash translation layer (FTL).” Among many FTL schemes available, the log block buffer scheme is considered to be optimum. With this scheme, a small number of log blocks, a kind of write buffer, can improve the performance of write operations by reducing the number of erase operations. However, this scheme can suffer from low space utilization of log blocks. In this paper, we show that there is much room for performance improvement in the log buffer block scheme, and propose an enhanced log block buffer scheme, called FAST (full associative sector translation). Our FAST scheme improves the space utilization of log blocks using fully-associative sector translations for the log block sectors. We also show empirically that our FAST scheme outperforms the pure log block buffer scheme.