Sharding is a widely adopted solution for blockchain scalability, yet its performance is often constrained by the high overhead of cross-shard transaction (CTX) processing. Existing systems typically rely on costly coordination protocols for atomic execution of CTXs, leading to high latency and reduced throughput. While recent advances in transaction allocation have minimized the proportion of CTXs, we argue that a critical aspect of sharding performance has been overlooked: the optimization of intra-shard transactions (ITXs), which now dominate the workload. In this paper, we introduce B-Shard, a novel sharding architecture that reorients the optimization objective toward maximizing ITX throughput, leading to significantly improved overall performance. The B-Shard design consists of a two-layer architecture: a global Beacon Shard for cross-shard coordination and multiple Worker Shards for transaction processing. By drawing inspiration from off-chain solutions, Worker Shards can rapidly process and finalize ITXs locally through batched commits to the Beacon Shard. The full potential of this architecture is revealed by a synergistic mechanism: our Balancing Transaction Scheduling algorithm (BATCH) buffers CTXs to create extended “pure ITX processing windows,” which are then processed with maximum efficiency by a lightweight intra-shard certification protocol. This synergy substantially reduces system overhead and maximizes throughput. Our evaluation against traditional twophase commit (2PC) and modern protocols demonstrates that B-Shard achieves a significant increase in total throughput. By accepting a minor latency increase for some CTXs, we achieve an overall reduction in average confirmation latency, validating the effectiveness of our strategy.

Pre-training based deep learning has been used for improving the accuracy of malicious traffic detection. When users use pre-trained encoders to avoid pre-training cost, they may suffer from backdoor attack. In other words, attackers can inject backdoors into these encoders, and trigger backdoors to evade malicious traffic detection. Thus it is necessary to detect the backdoors of these encoders before they are deployed. However, it is challenging, because: (1) traffic triggers are generated dynamically, (2) and encoders output embedding vectors. We hope to use the explanation collapse phenomenon (i.e., backdoor injection changes explanation results) to overcome these challenges. However, existing explanation algorithms are not suitable, because: (1) encoders output embedding vectors, (2) and encoders have no detection boundary. Thus it is necessary to design a special explanation algorithm. We propose MEB, a backdoor detection framework for pre-training based malicious traffic detection. In the detection framework of MEB, we use detection datasets and the explanation algorithm to calculate detection features, and input them into meta classifiers. We detect backdoors according to the prediction results of meta classifiers. In the explanation algorithm of MEB, we calculate the distance between the embedding vectors of explained samples and the embedding vector centroids of benign/malicious samples. We use the gradients of explained samples to the distance to calculate the explanation results. We implement and evaluate MEB in two typical malicious traffic datasets and two typical pre-training algorithms. Experiment results show that MEB can achieve higher detection accuracy and efficiency than baseline detection algorithms.

Money laundering using cryptocurrency poses significant threats to the blockchain ecosystem. Due to the decentralized and anonymous nature of cryptocurrencies, detecting such laundering activities is difficult. Although substantial research has been conducted, almost all existing methods detect cryptocurrency laundering from an individual perspective, ignoring the fact that money laundering is typically a group behavior. Group information should be very helpful in laundering behavior analysis, but such laundering groups are hard to be recognized due to anonymity and diversity of purposes of cryptocurrency transactions. To address this challenge, we design a multi-persona grouping algorithm that can effectively group accounts into persona subgraphs. Then, we extract two subgraph features: cycle basis number and cycle overlapping ratio, and build an unsupervised model to evaluate laundering scores of each subgraph. Extensive experiments on both synthetic and real-world datasets demonstrate that, compared with existing methods, our proposed method can improve detection accuracy by 17.4 percentage points on average. To the best of our knowledge, this is the first work on group-based detection of cryptocurrency laundering.

Communication cost is a main challenge in Federated Learning (FL). Gradient sparsification is one of the effective ways to reduce communication data volumes by allowing clients to send only a small portion of gradient elements to the server. Existing gradient sparsification methods, such as Top-K, analyze and decide the contribution of gradient elements for uploading by comparing their values. However, gradient elements with large values may not necessarily contain more information. To improve the accuracy while effectively reducing communication costs, in this paper, we propose a novel gradient sparsification-based FL algorithm called FedLOC, which sparsifies gradients not only based on the value of the gradient elements but also on the layer output of the local model. In particular, we design a two-phase compression operation to reduce the communication data volumes. First, FedLOC sparsifies gradients based on their values to construct initial compressed gradients. Subsequently, since the performance of the local model is related to the output of each layer, FedLOC analyzes the contribution of initial compressed gradient elements and constructs masks for the second sparsification operation based on each layer’s extend output derived from the original layer output. Convergence analysis and experimental results show that FedLOC can achieve superior performance in terms of accuracy and convergence when effectively reducing data communication volumes.

Scalability is one of the main problems limiting blockchain applications. Most recent research on blockchain scalability has focused on improving the consensus layer, but the most advanced consensus protocols can only reach a few thousand transactions per second(tps), which is far below the capacity of typical distributed databases. With the use of blockchain sharding technology, the time overhead corresponding to the execution layer and the consensus layer in the blockchain is gradually reduced. Additionally, improving the efficiency of transaction execution could incentivize nodes to actively verify transactions and enhance blockchain security. In this paper, we propose a Sharding Based Parallel Execution of Block Transactions (SPEx-Tran) framework in blockchain, which transforms the sequential execution of intra-block transactions into parallel execution to improve the performance of blockchain. In addition, based on the characteristics of blockchain sharding transactions, we further propose a Blockchain Transaction Scheduling (BCTS) algorithm to improve the performance of blockchain by optimizing the parallel execution process of transactions. In the BCTS algorithm, the mechanism of local multi-processors sharing the Conflict Transactions cache Table (CTT) simplifies the verification process of conflicting transactions, thus improving the blockchain performance. Experimental results show that the throughput of blockchain using the SPExTran framework and BCTS algorithm is significantly higher than that of typical sharding technologies.

Function as a Service (FaaS) has become a popular computing paradigm. FaaS shifts the burden of resource management to cloud providers and enables users to focus only on the logic of the code. At the same time, FaaS has thrived due to its scalability and billing strategies. However, cold start has been always a major challenge for FaaS which causes additional delay. Existing cold start mitigation methods ignore dependencies among functions resulting in bloated function combinations and consequently unnecessary resource consumption. In this paper, we propose FaaSCom, a FaaS scheduler which mitigates cold start problem from the perspective of preallocating containers. We mine causality among functions and divide the function dependency graph into communities. By scheduling containers according to the function communities, we can reduce the number of containers provisioned for functions to decrease resource consumption while not increasing cold start occurrences. Besides, we design an extended histogram policy to mitigate cold starts for long-time interval functions. We evaluate our method on an industrial serverless dataset. Compared with baseline methods, FaaSCom reduces the cold start rate by 8.4% while saving 50% memory consumption.

With the development of intelligent transportation system, it is urgent to transmit and analyze traffic data based on Internet of Vehicles. Federated learning has the characteristics of privacy protection, distributed learning and making full use of the computing power of local devices, which is very suitable for IoV's streaming big data analysis. Classical federated learning has only one global model, thus it can't be applied directly to IoV, where different region may have different data distribution. It is preferable that model of different region can be trained separately. Besides, the high mobility of vehicle nodes allows them to enter other areas while collecting data. This paper tries to solve the above issues. Firstly, a new vehicle-edge-cloud three-layer federated learning architecture for Io V, the Mobile Vehicle Multiple Access Federated Online Learning (MVMAFOL) algorithm, is proposed, where each vehicle terminal can upload the local model to multiple edge servers to maximize the use of its local model and effectively improve the generalization ability of the global model. Besides, a Model Similarity Vehicle Multiple Access (MSVMA) algorithm is proposed to improve the performance of MVMAFOL in the regional model aggregation phase, and an Edge Cloud Parameter Amendment (ECPA) algorithm is proposed for the global model aggregation phase. Simulation results show that this scheme can effectively improve the accuracy of the trained model.

Federated Learning (FL) has attracted a lot of attention from both academia and industry. In FL, user data no more need to be transmitted to the data center and each client device trains the deep model using its personal data so as to protect user privacy from being revealed. There exists two kinds of FL architecture, cloud based and edge based. Researchers have proposed the client-edge-cloud hierarchical FL system combining their advantage together to take the full advantage and avoid the defects. To improve the utilization ratio of local model parameters and data, so that enhance every single edge model accuracy, and ultimately achieve a more outstanding global model, we propose an algorithm Client Model Multiple Access (CMMA). CMMA allows clients associate with a set of edge servers, and uploads its training results to all servers in the set. That said, the local model of one client is reused by multiple edge servers. Such reuse can improve model performance particularly when the number of clients is small. Empirical experiments demonstrate the superiority of our scheme in different datasets, CNN models and data distributions. The results have not just shown the general suppress CMMA against hierarchical FL, but also validate the advantage of CMMA guaranteeing global model accuracy in unstable network or short of clients dilemma.

Distributed parallel training using computing clusters is desirable for large scale deep neural networks. One of the key challenges in distributed training is the communication cost for exchanging information, such as stochastic gradients, among training nodes. Recently, gradient sparsification techniques have been proposed to reduce the amount of data exchanged and thus alleviate the network overhead. However, most existing gradient sparsification approaches consider only synchronous parallelism and cannot be applied in asynchronous distributed training.