Continual event extraction is a practical task in natural language processing that requires models to learn quickly from new event types and data sources without forgetting pre-existing knowledge. It is important since language models often encounter new information to which they must efficiently adapt. Current work on continual event extraction either reuses existing parameters to learn from new tasks or statically assigns parameters specific to each new incoming task, incurring significant computational cost while preventing the potential knowledge sharing between tasks. To achieve a balance between the two, in this work, we present a means to adapt the model to incoming tasks in a parameter-efficient manner. We also incorporate metric learning to construct a prototypical network for maximum parameter efficiency. Experimental results on the ACE2005 dataset show that our framework maintains baseline performance with significantly smaller parameter sizes.
For event causality identification(ECI), many existing methods enhance the semantic information representation of inputs by using external knowledge or syntactic structures, while neglecting the mutual influence between event mentions. In this paper, we propose a new Mutual Attention based on Dependency Parsing(MADP) method, which combines the attention mechanism and the adjacency matrix obtained from the dependency parsing to enhance the connection between event mentions. Specifically, we design an attention mechanism, which includes trigger-aware attention and self-attention, to strengthen the connections between event mentions. The adjacency matrix helps our model eliminate the interference of redundant information on semantic information fusion. We evaluate our model on the dataset EventStoryLine(ESL) and achieve an F1 value of 65.1% , which outperforms the state-of-the-art model.
As Japan’s population continues to age, the rise in dementia cases has become a major societal issue. For early diagnosis and rehabilitation of dementia, research has focused on techniques that measure hand-motor skills to evaluate cognitive function. Previous systems faced challenges such as the need for improved hand movement evaluation methods, difficulty in system operation and implementation by end-users, and inadequate methods for data provision and visualization. This paper aims to enhance a web service to assess manual dexterity through tapping tasks. The proposed system has revised the evaluation method for scenarios where multiple tapping actions are performed under the same instruction, enabling more accurate assessments of motor function. Additionally, by employing Docker container technology to simplify system startup and converting recorded data into the more user-friendly XLSX format for use in applications like Excel, the system has improved data handling, making it more accessible for end-users.
The development of large language models (LLMs) has led to improved predictive performance and efficiency in various fields. However, the datasets used for training LLMs contain country-specific sentiment biases, which could pose serious issues in the field of security. This study aims to analyze the country-specific sentiment biases in the datasets used for training LLMs and evaluate their impact. In this study, we analyzed the sentiment tendencies associated with country names using VADER across four major large-scale training datasets: C4, RedPajama, OSCAR, and RefinedWeb. Specifically, we calculated the frequency of appearance and sentiment tendencies for each country, and further assessed the sentiment tendencies in bilateral relationships. The results revealed that the USA had a significantly high frequency of appearances and exhibited greater robustness compared to other countries. In contrast, negative sentiments towards Russia and Iran were prominently emphasized. Additionally, notable sentiment biases were confirmed in specific combinations of bilateral relationships. These findings indicate the presence of substantial biases in the datasets used for training LLMs. Particularly in the field of security, there is a risk of misunderstandings and conflicts arising in international relations. To address this, it is necessary to create datasets that evaluate LLMs’ perceptions of national conflicts and establish bias-removal standards in collaboration with international organizations. This will facilitate the development of more fair and reliable AI systems.
In parallel and distributed computing, there are practically two networks: linear networks (also called paths) and rings (also called loops). Many efficient algorithms, such as signal and image processing, were first discovered by solving algebraic problems, graphical problems and parallel implementations involving linear networks and rings. As a result, a network having both good path and cycle embedding is crucial. In this paper, using embedding method, we simulate the cartesian product of ( K_9 - C_9)^n graph into certain necklace graphs.
In beyond 5G (B5G) Internet of Things (IoT) system based on edge intelligence, pay-for-use demand has become a consensus, and the pricing of IoT services has attracted the attention of academia and industry. The pricing method based on noncooperative game allows edge service providers (ESPs) to compete fairly, effectively preventing edge nodes from malicious bidding. However, since only one winner can make a profit each time, it is easy to cause a large number of ESPs to lose money for a long time. To this end, a dynamic alliance pricing mechanism based on distributed edge intelligence is proposed. ESPs can freely choose to form an edge dynamic alliance, which not only retains the independence of edge nodes but also makes full use of the advantages of mutual cooperation between nodes. According to the characteristics of edge nodes, various roles are reasonably divided. In order to prevent abnormal behaviors of edge nodes, we set up necessary restrictive rules. At the same time, we designed a privacy-enhanced joint pricing prediction algorithm to screen the alliance's candidate solutions to improve pricing efficiency and edge benefit. The experimental results show that, compared with the traditional alliance game method, the performance of the mechanism we proposed improves the utilization rate of edge resources by 32.76%-61.37%. Meanwhile, the prediction accuracy was improved by 16.47%-38.86%, and the average prediction time was reduced by 42.81%-65.57%.
Binary code similarity detection (BinSD) systems trend to utilize deep learning to identify semantic features of assembly code and exhibits superior performance, gaining increasing popularity against traditional methods. However, it has been observed that existing deep learning models are susceptible to data poisoning attacks, posing a latent threat to the robustness and reliability of BinSD. Existing data poisoning strategies in BinSD are easily detectable for the generated triggers will destroy code functions. Moreover, selecting trigger injection location needs repeated exploration and verification, increasing the attack cost. To address this issue, we propose a novel adversarial scheme, named as AdvBinSD, which can poison the deep learning-based binary code similarity detector and make it sensitive to isolated instruction sequences. In AdvBinSD, the isolated instruction sequences generally refer to those instructions that have no data dependencies with other instructions and do not affect the function of original binary code, and also it is difficult to discovery those isolated instruction sequences by verifying syntactic validity and semantic integrity. Different from existing data poisoning strategies, AdvBinSD first estimates a code fragment that has the greatest impact on software functionality as the poisoning location, and then add isolated instruction sequences into this location to synthesize effective poisoned samples. This location estimation is achieved by maximizing the similarity between function-level feature vectors and instruction-level feature vectors, ensuring that the modified assembly code can execute correctly. Furthermore, to improve the efficiency of feature vector similarity computing process, a k-order greedy feature comparison (k-GFC) algorithm is also designated. Extensive experiments demonstrate that our proposed AdvBinSD can successfully poison the state-of-the-art deep learning-based binary code similarity detectors.
Enterprise networks are divided into sub-domain structures for resource aggregation and trust dependency. However, the complex domain structure results in delayed and incomplete defenses from Intrusion Response Systems(IRS). This paper proposes a proactive intrusion response system specifically designed for multi-domain networks called ProMD. A Bayesian Risk Graph is designed to analyze and evaluate the risks dynamically. The paper formalizes the selection of countermeasures as a multi-objective optimization problem (MOOP) and proves it is NP-Hard. Therefore a genetic-based heuristic algorithm OPGA is designed to generate optimal response policy sets and an unique policy is selected based on the Entropy Weight method. Experimental results demonstrate that the OPGA outperforms benchmark algorithms.
Bit-Flip Attack (BFA), an adversarial attack on network parameters, has proven to be a great threat to the security of Deep neural networks (DNNs). More concerning, recent works introduce a variant of BFA called Targeted Bit-Flip Attack (T-BFA), a targeted and stealthy attack paradigm. DNNs under T-BFA misclassify targeted inputs and behave normally on untargeted inputs, which makes the attack undetectable for users. Although existing defense methods can mitigate the effect of T-BFA, their lack of consideration for T-BFA characteristics results in unnecessary accuracy loss or run-time overhead. In this study, we leverage the “targeted” characteristic of T-BFA and introduce a two-step defense method to enhance the robustness of the DNN against T-BFA more cost-effectively. Firstly, in order to protect the security-sensitive classes that we identified based on T-BFA preferences from misclassification, we propose a method to locate security-sensitive weights crucial for classification of security-sensitive classes in convolutional and fully-connected layers. Secondly, in order to mitigate the effect of T-BFA and achieve a better trade-off between robustness and accuracy. We propose Random Switch to reduce the expected impact of bit flips on weights, and a selection strategy based on the benefit and negative impact of protecting a weight with Random Switch to further select security-sensitive weights for deployment of Random Switch. Experimental results show that our proposed method effectively enhances the robustness of DNNs against TBFA. Compared with the baseline, our method can tolerate more than $5.1 \times $ flipped bits with negligible accuracy loss (< 0.1% in resnet-50).
In various applications in Internet of Things like industrial monitoring, large amounts of floating-point time series data are generated at an unprecedented rate. Efficient compression algorithms can effectively reduce the size of data, enhance transmission performance and storage efficiency, and simultaneously lower storage costs. Therefore, there is a need for lightweight and efficient stream compression algorithms. In this paper, we propose a novel lossless floating-point data compression algorithm called Ant. The main idea is to encode double-precision floating-point numbers into integer form, calculate the delta between adjacent values, and then convert the delta into unsigned integers. This encoding method effectively reduces storage costs and improves data compression efficiency. Extensive experiments on real-world datasets demonstrate that our algorithm achieves compression speeds at least as fast as state-of-the-art streaming methods, and a 63% relative improvement in average compression rate.
Location-Based Skyline Query (LSQ) is a computational method aimed at retrieving non-dominated data points from spatial datasets. As data owners have limited resources, it is common for them to outsource data to clouds and provide query services for others. But this raises privacy challenges. Existing privacy-preserving LSQ schemes lack practicality due to computational or spatial overheads, with some requiring the data owner to perform heavy computations. To address this, we propose an innovative privacy-preserving LSQ scheme ensuring both data owner and query user privacy. We design a privacy-preserving point-convex intersection determination scheme (PPID) that enables a cloud server to determine whether a point intersects a polygon without knowing the real data. Based on PPID, we devise an efficient privacy-preserving LSQ scheme (PLSQ). Security analysis demonstrates PPID and PLSQ are safe in real/ideal world model security, and PLSQ can preserve dataset and query privacy. Extensive experiments reveal PLSQ’s efficiency, achieving search speed enhancements of $4 - 7 \times $ and decreasing data owner overhead by 25% – 40% compared to related schemes.
Vehicle-to-vehicle (V2V) energy trading stands as a significant technology, allowing electric vehicles (EVs) to share energy. This balances energy demand and supply, reducing pressure on the power grid. However, two substantial challenges require attention: 1) Creating a standardized protocol universally adoptable by EV manufacturers and energy trading participants to establish a common secure and reliable platform for real-time energy trading between EVs. 2) Determining a fair and transparent pricing and transaction management mechanism, accounting for factors such as energy demand and supply, location, and user preferences. To tackle these hurdles, this paper proposes a decentralized V2V energy trading system (VETS) based on a novel high-performance sharding blockchain framework. VETS incorporates a standard energy trading management protocol, achieving trustless and transparent transaction processing. It employs a blockchain sharding and parallel architecture for high transaction throughput, catering to a large number of EVs. Furthermore, an efficient consensus algorithm is developed to minimize service delays and adapt to real-time transaction demands. A practical testbed operationalizes VETS, facilitating experimental analysis within a real-world cloud environment.
This paper proposes a talking face generation method named "CP-EB" that takes an audio signal as input and a person image as reference, to synthesize a photo-realistic people talking video with head poses controlled by a short video clip and proper eye blinking embedding. It's noted that not only the head pose but also eye blinking are both important aspects for deep fake detection. The implicit control of poses by video has already achieved by the state-of-art work. According to recent research, eye blinking has weak correlation with input audio which means eye blinks extraction from audio and generation are possible. Hence, we propose a GAN-based architecture to extract eye blink feature from input audio and reference video respectively and employ contrastive training between them, then embed it into the concatenated features of identity and poses to generate talking face images. Experimental results show that the proposed method can generate photo-realistic talking face with synchronous lips motions, natural head poses and blinking eyes.
The rapid development of intelligent transportation technology has promoted the progress of multiple trains cooperative technology. This paper proposes an online cooperative cruise control method based on improved particle swarm optimization (IPSO) and deep feedforward neural network (DFNN), which enables each train to adaptively adjust the controller according to its own state information, ensuring speed synchronization and maintaining safe distance intervals between trains. Through the collection of information from the train’s big data, the optimal time-varying cooperative control gain is fitted to achieve efficient, safe, and real-time control of multiple high-speed trains.
With the application of blockchain light nodes in embedded devices, how to alleviate computing pressure brought by complex operations such as transaction’s SPV Verification for CPU of embedded devices and improve the performance of devices in these aspects has gradually become a research topic in industry and academia. This paper proposes a series of methods to improve the performance of blockchain SPV Verification from the perspectives of system architecture and hash computing unit: (1) According to the computational characteristics of SPV Verification, this paper customizes macro instructions and microinstructions for the coprocessor to meet the requirements of flexibility; The built-in dedicated cache holds transaction data fetched from external memory and intermediate data generated by internal Hash Computing Unit, which not only prepares transaction data for hash computation, but also avoids frequent access to the bus and external memory. (2) Techniques like two-round unfolded computing, timing-balanced pipeline architecture and optimized adders are adopted to improve the performance of SHA256 computation. (3) When double hash computing is required for transactions, Hash Computing Unit can directly perform the second hash computation based on the first hash computation, reducing the frequency of accessing to external memory, thereby improving the performance to a certain extent. Through these methods, the performance of the hardware coprocessor for SVP verification of transactions is more than double that of traditional solutions.
In the context of urban road following scenarios, ensuring the precise speed prediction and control of electric vehicle while rigorously maintaining the safe driving conditions has persistently presented a formidable challenge. To address such issues, a novel data-driven approach for the energy efficient speed planning of electric vehicles in vehicle-following scenarios is proposed. Incorporating Koopman operator theory, the developed control-oriented vehicle dynamics prediction model leverages real-world vehicle-following data in various vehicle-following scenarios. Subsequently, a data-driven Koopman model predictive control (MPC) method is developed based on the established prediction model, which focuses on reducing the energy consumption through speed planning while ensuring safe following distances. The comparative simulation results ultimately demonstrate the effectiveness of the proposed method under various driving scenarios.
The emergence of many-core processors presents significant opportunities for large-scale multithreading. Exploiting these intensive computing resources poses an urgent challenge for data processing systems. Although current big data frameworks offer automatic parallel computing capabilities, their coarse-grained parallelism hinders multi-threaded scalability on many-core processors. The NUMA architecture exacerbates load imbalance and resource wastage during the processing of skewed datasets. This paper introduces FinD, a fine-grained data processing framework that maximizes the utilization of many-core resources with lightweight threads and effectively addresses data skew through dynamic task scheduling. FinD incorporates bilateral NUMA-aware work stealing to achieve load balancing and reduce remote memory access across NUMA nodes. Experimental results demonstrate that FinD achieves a speedup of up to $9.7 \times$ compared to popular frameworks on big data benchmarks and HPC workloads while reducing overall execution time by 40% on skewed datasets.
In cellular networks, the video streaming services have been suffered from the limited network resources. For improving the video streaming services, decoupled uplink-downlink (DUDe) and multicast transmission with non-orthogonal multiple access (NOMA) are introduced to improve transmission efficiency. DUDe is considered beneficial to heterogeneous networks, which enable users to associate with different base stations (BSs) in uplink and downlink, respectively. Multicast transmission with NOMA is introduced to exploit the potential of multicast transmission, which enhances the efficiency of resource utilization. Combining the advantages of DUDe and multicast transmission with MOMA, we propose an auction mechanism to efficiently allocate resources and maximize the social welfare. However, the complex relationship between user value and resource configurations presents nonlinearity and non-convexity in the optimization problem. When calculating the user value in video streaming, most existing works do not consider the differences in individual user preferences. To address the problem, deep learning is utilized to model user value and transform the social welfare optimization problem into a mixed integer programming problem for solving efficiently. Our experiments show that the proposed mechanism is effective in heterogeneous networks.
In the field of mobile edge computing, the k Edge Server Placement (kESP) problem has attracted extensive attention. However, existing studies have mainly focused on solving small-scale problems, and the large-scale scenarios need to be explored. Based on this observation, this paper attempts to develop a cost-effective solution for the large-scale kESP problem. The proposed algorithm named UCIS-kESP realizes the selection of a small subset from a large collection of base stations by clustering mobile users, thereby significantly reducing the search cost of the kESP problem. Subsequently, an intelligent search based on genetic algorithm is applied to find the “optimal” deployment of k edge servers from the candidate subset. The experimental analysis conducted on a public dataset demonstrates that our proposed UCIS-kESP algorithm obtains better deployment solutions than the state-of-the-art algorithms, while also exhibiting fast and stable solving efficiency. On average, our algorithm can save about 40% of the computational time. Even as the size of the kESP problem increases, the computational overhead of the UCIS-kESP algorithm grows relatively slowly.
This study centers on the application of vertical federated learning technology in the context of Internet banking loans, with a particular focus on innovations in data privacy protection, risk control model algorithms, and secure multi-party computation. Currently, banking risk control strategies mainly rely on traditional data processing technologies, which often fall short in protecting user privacy and ensuring data usage efficiency. We adopt vertical federated learning technology, offering an innovative solution for the Internet banking loan scenario. Firstly, regarding data privacy protection, we propose a differential privacy mechanism to safeguard user-sensitive data. Secondly, we innovatively apply risk control model algorithms, facilitating collaborative modeling across multiple Internet loan platforms through federated learning. Furthermore, we introduce secure multi-party computation technology to ensure the secure transmission of data and confidentiality of computation processes during federated learning. Through empirical experiments on real Internet loan datasets, we validate the effectiveness and feasibility of our proposed methods. After implementing our risk control model, the credit approval rate increased from 3.44% to 18.2%, with a single-day high reaching 25.53%. The average loan amount increased by 7,700 yuan, and the average interest rate slightly declined by 0.48%, marking a significant improvement and breakthrough compared to traditional risk control models. This study offers innovative solutions for data privacy protection and risk control in the Internet loan scenario, providing safer and more reliable services for financial institutions and users. Moreover, our methods possess high practicality and promotional value. The potential widespread impact on the industry is profound.
