Home Archive
Archive
2026 Volume 63 Issue 11  Published: 2026-11-01
    Computer Architecture
  • Xuchuan Luo , Jiacheng Shen , Yangfan Zhou , Xin Wang
    doi: 10.7544/issn1000-1239.202550769

    Disaggregated memory (DM), as an emerging hardware architecture, has attracted extensive attention from both academia and industry in recent years. This architecture physically decouples computing and memory resources in data centers and interconnects them via high-speed networks, thereby improving resource utilization, system elasticity, and fault isolation capabilities. Distributed indexes, as the core component of memory-disaggregated storage systems, play a crucial role in efficiently organizing and maintaining data to support high-throughput, low-latency data access. However, due to the inherent resource constraints of DM, fully exploiting the benefits of resource disaggregation to maximize index performance remains challenging. A comprehensive survey of the research progress on memory-disaggregated indexing systems is presented. First, the fundamental features of DM and remote direct memory access (RDMA) are introduced, and the primary challenges in constructing high-performance indexes on DM are identified: concurrency control, read and write amplifications, and caching overhead. Subsequently, existing indexing systems are categorized into three types based on their underlying data structures: Hash indexes, tree indexes, and hybrid indexes. Representative systems in each category are analyzed to illustrate their design principles, implementation challenges, and key techniques. Finally, the main issues that require focused investigation in future research are highlighted, including cross-node cooperation, fault tolerance mechanisms, and adaptation to emerging software and hardware.

  • Computer Architecture
  • Yankang Zhao , Yuan Li , Xingyun Qi , Yihang Lu , Puguang Liu , Qiang Wang , Min Xie , Lin Peng , Mingche Lai
    doi: 10.7544/issn1000-1239.202550866

    With the rapid growth of neural network model sizes, single FPGA resources have become insufficient for accelerating large-scale models, while FPGA clusters face critical challenges, including complex architecture design, inefficient task partitioning, and the lack of automated toolchains. To address these issues, an end-to-end automated deployment framework for neural networks on FPGA clusters, named TH-FCAD (automated deployment framework for large-scale neural networks on FPGA clusters) is proposed. First, a two-level interconnect FPGA cluster architecture based on TH-Express is constructed. By integrating a high-speed global GTH (gigabit transceiver high-speed) network with a local LVDS (low voltage differential signaling) full-mesh interconnect, it achieves high-bandwidth and low-latency cross-node communication. Second, a latency-balancing optimization and model partitioning method is introduced to evenly partition large-scale neural networks into multiple acceleration kernels while maintaining inter-kernel delay equilibrium, which is essential for pipeline efficiency. Finally, a cluster-aware mapping algorithm is designed to automatically generate optimal placement schemes by jointly considering FPGA resource constraints and the cluster communication topology, thereby minimizing communication overhead. Furthermore, a complete automated toolchain is developed to streamline the entire deployment process, enabling seamless compilation from PyTorch/TensorFlow trained models to FPGA cluster configurations. Evaluations on a built 16 node FPGA cluster demonstrate that when deploying ResNet-8, SqueezeNet, and ResNet-50, TH-FCAD achieves inference latencies of 0.051 ms, 1.24 ms, and 1.36 ms, respectively, alongside a peak computational performance of 5.95 TOPS. Compared with state-of-the-art multi-FPGA approaches, the proposed framework delivers a 28.4% performance improvement and outperforms corresponding CPU and GPU implementations.

  • Computer Architecture
  • Wenbin Jiang , Baoyang Liu , Yukang Dong , Xinhai Shen
    doi: 10.7544/issn1000-1239.202550678

    In recent years, sparse-dense matrix multiplication (SpMM) has become a core operator in various fields such as scientific computing, graph processing, and deep learning, while also being a bottleneck limiting system performance. However, existing GPU implementations of SpMM generally suffer from low energy efficiency and resource utilization. Specialized accelerators demonstrate excellent energy efficiency in regular computations, while lack efficient support for irregular sparse matrix operations. To this end, we propose an efficient sparse matrix processing kernel tailored for dataflow architecture, DFU-SpMM, targeting the high-throughput dataflow unit (DFU). Firstly, a hierarchical sparse matrix processing method is proposed to ensure that the sparse blocks have a regular calculation pattern to reduce the number of dataflow graph retransmissions on the SpMM kernel on DFU. Meanwhile, for sparse blocks with different characteristics, different strategies are used to process them. Secondly, an adaptive assembly code generation method for dataflow graph optimization is proposed for optimizing dataflow graph. By leveraging the inspector-executor framework to overcome the current limitations of DFU in generating SpMM kernels, the dataflow graph is further optimized during code generation with respect to task allocation and register reuse. Experiments on representative datasets demonstrate that, compared with state-of-the-art implementations DTC-SpMM and HP-SpMM on an RTX 4090 GPU, the proposed method achieves an average 1.09 times and 1.23 times higher energy efficiency, along with 1.71 times and 1.10 times better computational resource utilization, respectively. Compared with the baseline SpMM kernel on the DFU accelerator without data reorganization, the optimized kernel with data reorganization achieves 1.41 times speedup and 1.51 times energy efficiency improvement.

  • Network and Information Security
  • Yiming Zhang , Fengwei Zhang , Shoumeng Yan
    doi: 10.7544/issn1000-1239.202550886

    With the rapid development of cloud computing, big data, and artificial intelligence, data security and privacy protection have become pressing challenges. In 2021, Arm proposed confidential compute architecture (CCA), which introduces hardware-level isolation mechanisms to provide a trusted execution environment for sensitive computing tasks, ensuring the confidentiality and integrity of both data and code during execution. Although initial progress has been made in CCA-based studies on user-level confidential execution, device isolation, and system-level security enhancement, the field as a whole remains in its early stages. Current research remains fragmented, with limited systematic analysis of architectural development and underlying technical frameworks. Moreover, some studies repeatedly reimplement fundamental mechanisms without generalizing common design, while others focus primarily on functional implementation with limited analysis of potential attack surfaces and long-term risks. To address these gaps, we systematically analyze the key mechanisms of CCA in confidential computing and review representative research advances in recent years. Specifically, we classify design strategies and implementation approaches in areas such as execution environment extension, device isolation, application adaptation, and security enhancement. Building on typical system case studies, we further summarize recurring technical patterns in CCA research, including root world isolation, multi-granule memory protection tables, and the coordination of stage-2 page tables with the system memory management unit. Finally, we discuss the major challenges and emerging trends faced by CCA, providing insights to guide future research and innovation in Arm-based confidential computing.

  • Network and Information Security
  • Siyu Chen , Mengdi Ouyang , Fagen Li
    doi: 10.7544/issn1000-1239.202550933

    In cloud computing environments, equality test can effectively address the challenge of comparing whether two ciphertexts are encrypted under the same plaintext without decryption. Among existing schemes, identity-based encryption with equality test (IBEET) significantly simplifies certificate management compared with traditional public-key encryption schemes, demonstrating considerable value. While China’s independently developed SM9 encryption algorithm has been successfully selected as an ISO/IEC international standard, existing SM9-based equality test schemes generally remain vulnerable to subversion attacks. Specifically, by subverting the randomness generation process, an adversary in possession of a backdoor key can recover the plaintext, thereby compromising the security of the scheme. To address this issue, an RCCA (relaxed chosen-ciphertext attack) secure identity-based SM9 encryption scheme with equality test (RCCA-SM9-IBEET) is proposed. The scheme employs cryptographic reverse firewalls (CRF) to re-randomize the inputs and outputs of potentially compromised users, thereby resisting subversion attacks. Furthermore, we define the concept of OW-ID-RCCA (one-way chosen-ciphertext security against a relaxed chosen identity attack) security model, which supports ciphertext re-randomization for compatibility with the reverse firewalls. A rigorous security proof is provided in the random oracle model. Finally, we analyze the computational and communication costs of the proposed scheme from both theoretical and experimental perspectives.

  • Network and Information Security
  • Tao Ye , Liuyang Jiao , Yongzhuang Wei , Lingchen Li , Erzhe Zhou
    doi: 10.7544/issn1000-1239.202550746

    Automated analysis techniques for lightweight block cipher algorithms have been a hot topic in recent research. The LLLWBC algorithm, proposed at the Inscrypt 2022 conference, is a novel low-latency lightweight block cipher that demonstrates excellent encryption and decryption performance in both software and hardware implementations. However, its security against emerging cryptanalytic techniques requires further investigation. In this paper, a new technique for constructing word-based propagation trails in block cipher algorithms is introduced, an automated detection method for seven distinct algebraic structures is proposed, then an automated search approach for integral distinguishers is also proposed based on these algebraic structures. Furthermore, the construction method for key guessing trails during the key recovery process of integral attacks is investigated, and the optimal key guessing set can be obtained by using this method. Based on the proposed method, we construct a 9-round integral distinguisher for LLLWBC with a data complexity of $ {2}^{52} $ plaintext blocks, which are one round longer than the distinguishers given by the designers. Subsequently, based on the obtained 9-round integral distinguisher, a 12-round key-recovery attack of LLLWBC is proposed, which requires the data complexity of about $ {2}^{56.17} $ chosen data, the memory complexity of about $ {2}^{72} $ 128b guess key and $ {2}^{52} $ 64b ciphertexts, the time complexity of about $ {2}^{122.09} $ times encryptions of 12-round LLLWBC. Experimental results demonstrate that the 12-round LLLWBC algorithm is vulnerable to integral attacks, while the full 21-round LLLWBC maintains a sufficient security margin. The construction method of integral distinguisher and key recovery technique proposed in this paper can also be applied to other word-oriented block ciphers.

  • Network and Information Security
  • Li Ma , Changjing Wang , Lin Zhang , Lixin Gan
    doi: 10.7544/issn1000-1239.202550653

    To prevent malicious access to and misuse of privacy-sensitive data, data collectors must ensure users’ explicit consent and autonomous control over the processing of their personal data, as mandated by modern data protection regulations. In existing approaches, users’ consent is mostly bound directly to data processing purposes. However, existing approaches suffer from vague semantics in consent derivation, weak alignment with the operational logic of data processing, and a lack of effective conflict resolution mechanisms, making it difficult to obtain legally valid and practically enforceable consent. To address these issues, a consent-based access control model is proposed, in which consent permissions are incorporated into the access decision process, and all personal data operations are abstracted as actions serving specific purposes. At the semantic level, the relationship between purposes and data operations is clarified, and an integrated semantic linkage among purpose, action, and consent is established. Following the principles of least privilege and deny-by-default, an asymmetric consent inheritance rule and a consent computation method tailored to hierarchical purpose structures are defined, thereby enabling fine-grained and context-aware access decisions. The model is formally specified in TLA+ (temporal logic of actions plus), and its core security properties are verified through model checking, ensuring both soundness and consistency. Experimental results on representative consent scenarios demonstrate that, compared with existing representative methods, the proposed approach can effectively resolve consent conflicts, reduce unintended over-authorization, and achieve better decision quality and compliance.

  • Network and Information Security
  • Yimin Guo , Qi Jiang , Ping Xiong
    doi: 10.7544/issn1000-1239.202550862

    The Internet of things (IoT) environment, characterized by resource-constrained terminals and sophisticated threats, poses severe challenges to the design of authentication protocols. Notably, even protocols verified by formal tools often remain vulnerable to implicit attacks, where adversaries combine multiple capabilities such as channel control and secret extraction, to launch more complex and concealed attack strategies. To address this issue, we propose a provably secure, resource-efficient, and lightweight authentication protocol. The protocol employs lightweight cryptographic primitives, including Hash functions and physical unclonable functions (PUFs), and optimizes both communication and computational overhead through dynamic identity and message fusion mechanisms, without relying on expensive public-key operations. To establish a solid security foundation, we construct an extended real-or-random (ROR) model capable of characterizing composite attack scenarios, under which the semantic security of the session key is rigorously proven. Heuristic analysis further demonstrates that the protocol achieves all nine predefined security properties even under implicit attacks. Performance evaluation and comparative analysis indicate that the proposed protocol requires only 0.5835 ms of computational time on smart devices and incurs a communication overhead of 3168 b. While fulfilling comprehensive security objectives, its resource efficiency significantly outperforms existing schemes, offering a practical solution that balances security and lightweight design for IoT applications.

  • Network and Information Security
  • Jiaxin Zhao , Jingzhou Wang , Yu’e Sun , He Huang
    doi: 10.7544/issn1000-1239.202660004

    Thanks to the reconfigurable ability of optical switching technologies, data centers are able to adjust topologies at runtime to adapt to traffic dynamics generated by new technologies and applications. However, existing solutions for reconfigurable and traffic scheduling suffer from the high flexibility of configurations and traffic dynamics. Static solutions overlook practical constraints, fail to adapt to real-production environments. Meanwhile, dynamic solutions cannot guarantee global optimization, resulting in decrease in overall network performance. To bridge the gap, we first formalize the joint reconfiguration and traffic scheduling as the k-weight limited matching problem and propose an approximation algorithm for global updates based on priority queues and greedy iteration. Then, we divide the traffic dynamics into three different scenarios, each of which is addressed by an efficient local scheduling algorithm. To further improve network performance of both global update and local scheduling, we propose SYNC (static and dynamic network coordinator) through a periodic global optimization and graph structure-aware mechanism to achieve the trade-off between throughput and scheduling efficiency. We conduct simulation experiments with both real-world datasets and synthetic datasets generated by RMAT (recursive matrices). Experiments show that SYNC provides an efficient, stable and scalable scheduling solution for reconfigurable data center networks. Compared with existing solutions, SYNC reduces running time by 36.56%, and exhibits high stability at runtime.

  • Network and Information Security
  • Yongji Liu , Chenglong Li , Anwen Huang , Qiong Li , Yangxurui Liu , Xin Song
    doi: 10.7544/issn1000-1239.202550643

    With the rapid development of integrated space–air–ground networks and large-scale constellations, low earth orbit satellite networks (LEO-SN) together with unmanned aerial vehicles (UAV) constitute a hierarchical and heterogeneous aerial access/backhaul backbone, realizing three-dimensional “space–air–ground” complementarity and becoming a research hotspot for 6G. Owing to the high packet-loss rate, long latency, and pronounced dynamics of LEO-SN links, the Transmission Control Protocol (TCP) congestion-control mechanisms designed for terrestrial networks cannot maintain stable and efficient bandwidth utilization. Compared with traditional loss-based and delay-based algorithms, probe-based algorithms represented by BBR (bottleneck bandwidth and round-trip propagation time) are better suited to such complex environments; nevertheless, theoretical analysis and experiments show that, limited by its passive sensing strategy, BBR can suffer a sustained throughput drop during satellite handovers—the so-called “long trough effect”. To address this issue, we propose SHP-BBR, a TCP congestion-control mechanism based on satellite-handover prediction. SHP-BBR leverages historical satellite-handover records and round-trip-time data to build a time-series prediction model that forecasts upcoming handover events. Guided by these predictions, SHP-BBR proactively senses impending link-state changes, obtains accurate network-parameter estimates, and preserves bandwidth utilization. Experimental results demonstrate that SHP-BBR stably eliminates the long trough effect in LEO-SN environments, achieving an average throughput gain of 33%. Moreover, compared with conventional TCP congestion-control algorithms, SHP-BBR delivers efficient, well-balanced throughput and latency under dynamic network conditions.

  • Artificial Intelligence
  • Qianhao Ren , Ruiheng Liu , Yu Zhang
    doi: 10.7544/issn1000-1239.202550826

    Traditional Text-to-SQL tasks typically require large scale annotated data for model training, which incurs significant time and labor costs. Consequently, data scarcity remains a critical bottleneck, particularly in domain specific applications. While data augmentation based on large language model (LLM) has emerged as a promising solution to this issue, most existing methods heavily rely on the distribution of original datasets and often generate synthetic data that lack reliability and interpretability. Addressing these challenges, we propose SCoT-SQL, a novel chain-of-thought guided data synthesis framework. This method effectively integrates natural language instructions with symbolic knowledge, specifically detailed database schema and table content, to guide LLM, thereby significantly reducing semantic errors during the generation process. To guarantee the high quality of synthetic data, SCoT-SQL incorporates a rigorous filtering mechanism that leverages both model self consistency voting and verification based on execution. Additionally, the framework employs a self feedback mechanism driven by chain-of-thought reasoning to calibrate synthetic data. This process not only rectifies logical inconsistencies in SQL queries but also generates explicit step by step reasoning trajectories, enhancing the interpretability of the dataset. Comprehensive experiments on the KaggleDBQA and ScienceBenchmark datasets demonstrate the efficacy of the proposed method. SCoT-SQL achieves execution accuracy improvements of 6.5% and 3% respectively, significantly outperforming previous state of the art data synthesis baselines and proving its effectiveness in low resource and complex scenarios.

  • Artificial Intelligence
  • Xue Tan , Yi Zheng , Yunruo Zhang , Yiran Chen , Ping Chen , Xiangyang Xue
    doi: 10.7544/issn1000-1239.202550713

    Retrieval-augmented generation (RAG) systems extend the capability of large language models by incorporating external knowledge databases, enabling more accurate and up-to-date responses. However, this integration also introduces a novel privacy vulnerability: mapping attacks (MAs), which aim to infer whether a specific private fragment is indexed in the retrieval database and to uncover its retrieval behavior. Such attacks exploit both embedding-level signals and retrieval dynamics, posing a serious threat to sensitive data protection in RAG pipelines. Despite the growing awareness of privacy risks in RAG systems, there is currently no dedicated defense mechanism specifically designed to mitigate mapping attacks. We introduce AMRP-SIP, a dual-randomization framework that concurrently protects both embeddings of the documents and retrieval traces, while preserving state-of-the-art utility. AMRP-SIP comprises three lightweight stages. First, a random orthogonal projection compresses each query and document into a low-dimensional latent space, hiding raw embeddings and reducing downstream noise. Second, adaptive differential privacy injects cluster-adaptive Gaussian noise, ensuring (ε, δ) fragment-level privacy. Third, a score-dropout layer introduces randomness by perturbing similarity scores with noise and probabilistically dropping a portion of the retrieved documents with probability p, thereby obfuscating the retrieval trajectory. Experiments on Wiki-40B, PubMed-256, and IP-DataSet demonstrate that AMRP-SIP reduces the AUC of membership inference attack (MIA) from 0.75 to 0.27, while maintaining competitive performance in downstream RAG tasks.

  • Artificial Intelligence
  • Yingchun Yang , Jingbo Chen , Xuemei Peng , Dong Ji , Songran Liu
    doi: 10.7544/issn1000-1239.202550290

    The diffusion models have made significant achievements in the field of image generation, but they still face some challenges such as high computational cost, long inference latency, and poor image generation quality during image generation. Based on previous research, we propose an efficient adaptive diffusion model inference method based on the warmup-aware and noise reuse methods, which can optimize the performance of the diffusion models according to input conditions and solve the problems of slow inference speed and poor image generation quality. Specifically, our warmup-aware method can dynamically adjust the number of warmup time steps in the initial stage of the diffusion process according to the semantic complexity of the input text prompts to ensure efficient initialization. Subsequently, our noise reuse method dynamically determines the number of time steps for noise reuse in the iterative refinement stage, and reduces unnecessary computational overhead. And we achieve 2.66 times acceleration on 4 NVIDIA V100 GPUs, and compared with DistriFusion, our inference speed increases by 12.8% and the image generation quality improves by 5.03 times. These improvements indicate that our proposed method has significant effects in balancing inference efficiency and image generation quality, and provides a scalable solution for real-time image generation applications.

  • Artificial Intelligence
  • Yourui Han , Sailike Saya , Naibijiang Aikeremu , Weihua Meng , Youpeng Hu , Ziyuan Li , Xuequn Shang , Bolin Chen
    doi: 10.7544/issn1000-1239.202550749

    The identification of cancer driver genes is an important research area in oncology research, holding significant value for a deeper understanding of the mechanisms of cancer development and achieving precision medicine. Currently, most research methods primarily rely on graph data to capture the characteristics of driver genes. However, the evolution process of cancer development involves many complex interactions among various biomolecules, making it difficult to fully reveal its essence using information from only single modality. Therefore, how to efficiently integrate multi-modal biological data to comprehensively identify driver genes has become a critical issue that needs to be addressed. To this end, we propose a cancer driver gene identification method, titled as TGMNN (text-graph modal neural network), based on a combination of text and graph modality. This method builds upon the graph modality by further incorporating text semantic information related to genes and employs an interactive attention mechanism to achieve collaborative fusion of multi-modal data, thereby reducing the differences between modalities. Compared with existing methods, TGMNN can capture the information interactions more effectively between modalities, significantly improving the accuracy of driver gene predictions. Experimental results show that AUC value of this method reaches 0.897, outperforming existing methods and high-lighting the model’s effectiveness.

  • Artificial Intelligence
  • Peng Ying , Zhongnian Li , Hongguang Yuan , Meng Wei , Xinzheng Xu
    doi: 10.7544/issn1000-1239.202550731

    In recent years, large vision-language models (VLM) have demonstrated remarkable zero-shot prediction capabilities in image classification tasks. Existing VLM-based classification approaches typically rely on a set of predefined category labels or external textual knowledge bases as semantic prior to align and match visual content with language. However, in real-world applications, such predefined labels often fail to cover the vast and dynamic concept space of the open world, thereby limiting the adaptability and generalization ability of these models. To address this challenge, a more open-ended task, generative image classification (GIC), is introduced, in which no prior semantic labels are provided, and the target categories reside in an open-ended, potentially unbounded semantic space, substantially increasing the classification difficulty. To tackle the complexity of GIC, a granularity-guided multi-model collaboration framework is proposed, designed to fully leverage the strengths of VLM in both image understanding and text generation. In this framework, prompt templates are carefully crafted to guide VLM in generating candidate labels at varying levels of semantic granularity. Subsequently, the candidate label space is expanded by a language model that associates semantically related categories based on the generated image descriptions at each granularity level. Finally, the most relevant label is identified through a semantic similarity retrieval process in the embedding space. The effectiveness and superiority of the proposed method for the GIC task are validated through extensive experiments on multiple fine-grained image classification benchmarks.

  • Artificial Intelligence
  • Jianfeng Yang , Dongpo Zhu , Rui Cao , Xuanhao Cui , Ping Wang , Jie Zheng
    doi: 10.7544/issn1000-1239.202550936

    Existing image captioning methods often struggle to simultaneously capture global semantics and fine-grained details, leading to overly generic descriptions or missing key information. To address this issue, we propose EAFCC, short for expert-aware fusion with common-specific collaboration, a multi-feature image captioning model with mixture-of-experts fusion. The model takes grid features, region features, and segmentation features as parallel inputs. In the encoding stage, a pre-fusion module first performs implicit alignment of heterogeneous visual features and models cross-feature common information within a shared semantic space, while preserving the basis for feature-specific representations. A shared Transformer encoder is then employed to learn the common representations, whose consistency and discriminability are further enhanced by a contrastive learning constraint. Meanwhile, three independent encoding branches are used to model feature-specific representations, so as to retain fine-grained complementary information associated with different feature types. A gated routing mechanism inspired by mixture-of-experts is further introduced to dynamically weight and fuse the specific information. Finally, the enhanced common features and the fused specific features are collaboratively integrated, processed by a skip-connected encoder, and fed into a Transformer decoder to generate image captions. Experimental results on MS-COCO dataset show that EAFCC achieves competitive performance on the CIDEr-D metric and outperforms existing single-feature and dual-feature methods, demonstrating the effectiveness of common-specific collaborative modeling and dynamic multi-feature fusion for image captioning.

  • Artificial Intelligence
  • Yiqin Luo , Tianlong Gu , Fengrui Hao , Liang Chang
    doi: 10.7544/issn1000-1239.202550799

    Deep learning technology has been increasingly empowering the practice of dermatological disease diagnosis, leading to a steady improvement in its diagnostic performance. However, such data-driven deep learning models are plagued by severe bias issues. If systems with these biases are deployed in clinical applications, they are bound to cause unfair treatment towards specific groups, such as underrepresented groups. Ensuring the fairness of decision-making has become a critical and urgent problem that needs to be addressed. Existing research has indicated that enhancing fairness often comes at the cost of reduced diagnostic performance. Therefore, to achieve a better trade-off between diagnostic performance and fairness, while simultaneously alleviating the unfair treatment of underrepresented groups, we propose an enhanced balanced incremental distillation network (EBID-Net). Specifically, with the aid of balanced memory, the model leverages representative demographic groups to assist underrepresented groups in knowledge learning, while being incrementally trained by integrating distributions across different groups. Additionally, we incorporate global information into the contextual attention mechanism to capture correlated interactions between features across different spatial locations, thereby obtaining more robust feature representations during incremental learning. Experimental results show that EBID-Net outperforms other methods in fairness criteria and in the trade-off between fairness and diagnostic performance.