Abstract:Decompilation serves as a fundamental technique in software reverse engineering, aiming to recover high-level source code from hardware-oriented binary programs to support human understanding, analysis, and re-engineering tasks. Although this technique has been extensively studied, traditional rule-based decompilers often generate decompiled code with poor readability and limited reusability. Moreover, due to long development cycles, support for emerging instruction set architectures such as RISC-V is typically delayed in conventional decompilers. With the widespread adoption of large language models (LLMs) in automated software engineering tasks and their demonstrated effectiveness, this study proposes BinDec, a RISC-V binary decompilation approach that synergistically integrates LLM and symbolic execution. The proposed method alternates between LLM-based code generation and symbolic execution-based verification, fully exploiting the code understanding and generation capabilities of LLM to produce decompiled code that is more readable and reusable, while leveraging the analysis and verification capabilities of symbolic execution to ensure semantic correctness and reliability. The effectiveness of the proposed method is evaluated through a series of experiments. Experimental results demonstrate that BinDec achieves semantic accuracy comparable to that of traditional decompilers, while significantly improving the readability of the generated decompiled code.

Abstract:The RISC-V instruction set architecture, characterized by its openness and modular design, promotes innovation and customization in processor architecture, while simultaneously introducing severe software ecosystem fragmentation. Traditional cross-platform software build mechanisms, such as on-site compilation, IFUNC, and Multilib, encounter significant challenges in the RISC-V ecosystem, including limited compatibility, high maintenance overhead, and insufficient optimization granularity, which highlights the need for new solutions. To address these issues, this study proposes a compatibility-aware multi-level compilation method for the RISC-V platform—RuyiBuild toolchain. By adopting LLVM IR as the intermediate representation and integrating a non-intrusive compilation interception mechanism, transparent adaptation to existing software build systems is achieved. As a result, the generated operating system software packages simultaneously support compatibility across heterogeneous combinations of RISC-V extension instruction sets and adaptive, extension-aware optimizations tailored to target execution platforms. This approach systematically resolves the dual-objective challenge of achieving both high-performance binary optimization and broad compatibility under diverse instruction set extensions and platform variations inherent in the RISC-V ecosystem. Centered on the extraction, deployment, transformation, and optimization of LLVM IR, RuyiBuild establishes a comprehensive framework for cross-platform software distribution and fine-grained optimization. The framework consists of four core mechanisms: a transparent dual-path compilation and LLVM IR extraction mechanism; a dynamic library LLVM IR aggregation and link-time transformation mechanism; an LLVM IR deployment and automated RPM integration mechanism; a client-cloud collaborative LLVM IR dynamic transformation and resource-adaptive scheduling mechanism. From an implementation perspective, the RuyiBuild toolchain encapsulates compilation tools and system commands, enabling full-path LLVM IR extraction and distribution without modifying source code or existing build systems. During deployment, synchronized distribution of LLVM IR and traditional binaries is supported. Furthermore, to facilitate performance optimization across various target devices and microarchitectures, client-side resource-aware deferred transformation and cloud-side multi-architecture adaptation with dynamic LLVM IR transformation are provided. Experimental results show that RuyiBuild supports deployment across a wide range of RISC-V instruction set extensions and microarchitecture combinations, while achieving a favorable balance among performance, compatibility, build overhead, and deployment complexity. Consequently, this study provides a novel and effective solution for software build, deployment, and adaptation in the RISC-V ecosystem, offering both academic value and practical potential.

Abstract:With the rapid proliferation of the open-source RISC-V architecture, its openness and modular design foster a thriving hardware ecosystem while simultaneously posing significant challenges to hardware security assurance. Early detection of security vulnerabilities during the initial stages of the hardware design flow enables vulnerabilities to be eliminated at minimal cost before being permanently embedded into physical silicon. Although static analysis techniques are applied to early-stage hardware security vulnerability detection, existing methods suffer from high false negative rate (FNR) and false discovery rate (FDR) due to insufficient utilization of specification knowledge and an inadequate semantic understanding of code context. To address these challenges, this study proposes MARC, an early detection method for hardware security vulnerabilities based on collaborative multi-agent systems powered by large language models (LLMs). The proposed method constructs a collaborative framework consisting of four specialized agents: design dependency analysis, documentation analysis, security vulnerability detection, and security vulnerability confirmation. Through multi-dimensional collaboration, including cross-module context augmentation, structured utilization of module documentation knowledge, rapid preliminary screening of potential vulnerabilities, and deep reasoning-based risk analysis. Experiments conducted on an industrial-grade dataset demonstrate that the MARC framework achieves an FNR of 0.3829 and an FDR of 0.3695, representing reductions of approximately 18.2% and 19.1%, respectively, compared to the baseline method. The proposed framework effectively reduces false positives and improves detection accuracy in early-stage hardware design. Furthermore, its real-world effectiveness is validated by the discovery of a hardware vulnerability that has been assigned a CVE identifier. By its successful application, the authors’ team won the global championship in the HACK@DATE 2025 hardware security competition.

Abstract:With the rapid development of edge computing and intelligent end-side software, the RISC-V architecture attracts increasing attention in both academia and industry due to its open-source, modular, and low-cost characteristics. However, deploying intelligent software on the RISC-V architecture presents significant challenges. The execution of intelligent software relies on deep learning operators, such as convolution, matrix multiplication, and normalization. Once defects occur in these operators, the execution efficiency, accuracy, and reliability of a large number of upper-layer intelligent applications are directly affected, making operator quality evaluation critically important. Existing operator testing methods are primarily designed for x86 architectures and have difficulty characterizing the impact of varying computational complexity on RISC-V platforms under constraints such as limited memory capacity, power consumption, and system resources. To address this issue, this study proposes RIVdoo, a deep learning operator testing method tailored for the RISC-V architecture. RIVdoo systematically covers different computational workloads through a grouping strategy based on operator input-space complexity, evaluates accuracy, execution efficiency, memory usage, and system overhead using multidimensional metrics, and introduces a differential testing mechanism with complexity amplification factors to effectively identify performance anomalies and architectural adaptation issues. Experiments covering 47 operators across compute-intensive, memory-intensive, and lightweight categories demonstrate that RIVdoo reveals significant performance trade-offs and adaptation differences among existing operator libraries and optimization strategies on the RISC-V architecture. Specifically, TFLite trades 1.5×–2× memory overhead for 30%–50% performance improvement, which leads to memory bottlenecks in resource-constrained scenarios; TVM’s dynamic storage scheduling incurs 60%–150% higher page fault rates than TFLite due to the limited TLB and cache capacity of RISC-V; RVV vectorization causes performance degradation for certain operators and low-complexity workloads because of software emulation and startup overhead, indicating that existing optimization strategies lack targeted design for RISC-V platform characteristics. The results demonstrate that the runtime behavior of different operator implementations and optimization strategies on the RIVdoo architecture exhibits strong complexity-dependent characteristics, and that output correctness alone is insufficient to reflect real deployment performance. RISC-V provides a systematic evaluation methodology for operator adaptability analysis and optimization on RISC-V platforms.

Abstract:Image processing tasks are rapidly migrating to cloud and multi-party collaborative environments. However, directly processing plaintext image data on cloud servers easily leads to the leakage of sensitive information in images and is difficult to resist malicious attacks such as tampering, thus failing to guarantee data integrity and service reliability. To address these challenges, this study proposes MalOD, an object detection inference framework for environments with malicious adversaries. MalOD is a framework to achieve secure object detection under malicious adversaries. MalOD constructs an encrypted feature pyramid network (SecFPN) to perform multi-level feature extraction on encrypted images. Based on these multi-level cipher text features, a secure region proposal network (SecRPN) and a secure region of interest align (SecRoIA) module are designed to achieve secure object detection. By leveraging replicated secret sharing (RSS), a series of secure computation primitives are designed, including a secure ceiling function, secure bilinear interpolation, and secure nearest-neighbor interpolation. These primitives provide the underlying support for SecFPN, SecRPN, and SecRoIA, ensuring the efficiency and accuracy of the detection process under malicious adversaries. The correctness and security of MalOD are proved, and its performance is evaluated on the COCO 2017 and Pascal VOC 2012 datasets. Experimental results show that MalOD achieves high object detection accuracy while meeting strict security requirements. In particular, when the intersection over union (IoU) threshold is 0.5, the average precision on the COCO subset decreases by only 0.113 compared with plaintext detection. This study provides theoretical and practical support for privacy-preserving image processing under malicious environments and is particularly suitable for untrusted cloud computing and multi-party collaboration scenarios.

Abstract:Software programming assistants based on large language models (LLMs), such as Copilot, significantly enhance programmer productivity. However, LLMs have large computing and storage requirements and are difficult to deploy locally. Building a lightweight, small LLM can meet computing, storage, and deployment requirements, but it leads to a greater accuracy loss in code generation compared to large LLMs. Knowledge distillation (KD) techniques allow small LLMs (student models) to approximate the output distributions of large LLMs (teacher models) on target training datasets, thus reducing accuracy loss in code generation. Cutting-edge KD techniques in artificial intelligence are based on the Kullback-Leibler (KL) divergence loss function, which measures and reduces accuracy loss due to discrepancies in the output distributions between student and teacher models. However, student models struggle to learn in the near-zero distribution regions of teacher models. Consequently, researchers have employed the reverse KL (RKL) divergence loss function to address this issue in near-zero distribution regions. This study finds that RKL faces learning challenges in high-probability distribution regions and complements the KL divergence loss function. For some datasets, low-quality outputs from teacher models lead to poor learning outcomes for the student models. This study proposes an adaptive knowledge distillation (AKD) method that uses prompts to enhance teacher model output quality and constructs an adaptive loss function to adjust learning priorities based on the distributional differences between student and teacher models. This ensures the student model effectively learns in both primary and near-zero probability regions. Using the AKD method, this study trains a lightweight large code model based on StarCoder-1B/7B (student/teacher models) and the CodeAlpaca dataset, evaluating accuracy loss and code quality issues. Experimental results show that the lightweight model size is reduced by 85.7%. On the HumanEval and MBPP data sets, prompts with clear instructions improve teacher model code generation quality, reducing the average accuracy loss of the trained student model by 6%. The AKD-trained model’s average accuracy loss compared to the teacher model (StarCoder-7B) is 17.14%, a 30.6% reduction over the original student model. The AKD-trained model’s accuracy loss is reduced by an average of 19.9% compared to state-of-the-art KD and RKD methods. Regarding inference memory requirements, the KD and RKD methods require 54.7 GB, while the AKD method only adds 3 GB. In terms of training time, the AKD method incurs a 30% increase. However, even when the KD and RKD methods are trained for the same duration, their average performance improves by only 3%, which is 16.9% lower than that of the AKD method. Therefore, the additional training cost of the AKD method is justified. Moreover, applying the AKD method to the Code Llama and CodeGen model series reduces accuracy loss by an average of 19.2% compared to state-of-the-art KD and RKD methods, demonstrating the generalizability of the AKD method.

Abstract:Android application developers need to quickly and accurately reproduce error reports to ensure application quality. However, existing methods often rely solely on crash information provided in stack traces to generate event sequences, making it difficult to accurately locate the crash page and offer effective guidance for dynamic exploration to trigger the crash. To address this issue, this study proposes a component-aware automatic crash reproduction method for Android applications, called CReDroid, which effectively reproduces the crash by leveraging both the title and stack trace of the crash report. First, CReDroid dynamically explores the application under test to construct a component transition graph (CTG) and combines the dynamic exception information from the stack traces with the static component interaction data from the CTG to accurately locate the target crash component. Second, based on the critical operations in the crash report title and the reachable paths in the CTG, CReDroid designs an adaptive strategy that uses the contextual relationship between the current page’s component and the crash component to assign priority scores to GUI widgets. The dynamic exploration process is globally optimized through reinforcement learning to effectively reduce inaccuracies in the prediction process. This study evaluates CReDroid using 74 crash reports and compares its performance with state-of-the-art crash reproduction tools, including CrashTranslator, ReCDroid, and ReproBot, as well as widely used automated testing tools, Monkey and APE. The experimental results show that CReDroid successfully reproduces 57 crash reports, which is 13, 25, 27, 30, and 17 more than CrashTranslator, ReCDroid, ReproBot, Monkey, and APE, respectively. Moreover, for the successfully reproduced crashes, CReDroid reduces the average reproduction time by 26.71%, 94.96%, 71.65%, 84.72%, and 88.56%, compared to CrashTranslator, ReCDroid, ReproBot, Monkey, and APE.

Abstract:With the continuous development of blockchain technology and applications, the demand for interaction between blockchains is increasing. However, the lack of effective interoperability between different blockchain systems limits the further development of blockchain technology. To address the problem of heterogeneous interconnection between blockchains, cross-chain technology has emerged and quickly become a prominent research topic. Specifically, the XCMP protocol, one of the most popular cross-chain communication protocols, not only provides a secure and efficient communication mechanism but also offers a broad platform for future blockchain innovation and applications. However, the cross-chain massage passing (XCMP) protocol is still in a phase of continuous development and improvement, facing security challenges such as replay attacks, denial of service attacks, and delay attacks. This study formally verifies and improves the XCMP protocol, aiming to provide solid support for the development of more secure and feature-rich decentralized applications based on it. First, Z language, a formal description language based on classical set theory and first-order predicate logic, is used to summarize, refine, and formally model the 10 key security goals and protocol contents of the XCMP protocol. The security goals are then verified using Z/EVES, an automated verification tool supporting the Z language. The verification results show that the XCMP protocol does not meet three of the security goals. Second, after a comprehensive analysis of the verification results, the study introduces a commitment mechanism, a supervision mechanism, and a polling mechanism to address unmet security goals of the XCMP protocol, proposing an enhanced cross-chain message passing (E-XCMP) protocol. Finally, the E-XCMP protocol is formally modeled, and its security and reliability are evaluated using the security protocol analysis tool Scyther and the automatic verification tool Z/EVES. The evaluation results show that the E-XCMP protocol not only meets the three previously unmet security goals but also effectively solves security issues such as replay attacks, denial of service attacks, and delay attacks, demonstrating strong security and reliability.

Abstract:Existing static malware similarity measurement methods are affected by static anti-antivirus techniques, and the model features are either easily confused or fail to fully capture malware semantics. This study proposes a malware similarity measurement method called heterogeneous graph matching network-based similarity (HGMSim) to address the above problems. This method first uses the disassembly tool IDA Pro to extract a malware’s call graph, which is then abstracted into a heterogeneous graph to effectively capture the heterogeneous semantics of different function node types and their call relationships. Meanwhile, cross-graph edges are established for similar function nodes of the same type in two call graphs to mine the implicit neighbor semantics between nodes in different call graphs, and a heterogeneous graph matching network is constructed. Then, the study proposes a heterogeneous graph embedding method based on local node graph matching strategy and implements malware similarity measurement to solve the problem of difficulty in distinguishing malware with highly similar graph structures between different families. Finally, experimental results show that HGMSim performs best in malware similarity measurement.

Abstract:With the continuous advancement of compilation technology, modern compilers support richer programming models and more complex compilation optimizations, which makes manually adjusting compilation options for optimal performance extremely challenging. Although various automated compilation tuning methods have been proposed, traditional heuristic search algorithms often struggle to avoid being trapped in local optima when confronted with vast search spaces. Moreover, most existing tuning methods target single-core or multi-core architectures, limiting their use in large-scale parallel computing systems. To address these issues, this study designs and implements a distributed compilation tuning framework, SWTuner, based on machine learning methodologies. By introducing AUC-Bandit-based distributed meta-search strategies, machine learning model-guided performance prediction, and SHAP-based compilation option analysis and filtering, the resource utilization and search efficiency during the compilation tuning process are significantly improved. Experimental results show that SWTuner performs excellently in tuning typical test cases on the new-generation Sunway supercomputer, not only reducing search time but also achieving notable reductions in actual execution power consumption during the search process compared to other tuning methods. During the tuning process, the random forest model employed by SWTuner demonstrates good generalization capability and prediction accuracy, effectively reducing search space dimensionality while maintaining tuning effectiveness, providing an efficient and reliable solution for automatic compilation tuning in high-performance computing.

Abstract:Large language models demonstrate significantly superior performance in reasoning tasks compared to traditional models, yet still struggle to meet the demands of complex tasks in terms of computational cost and response quality. Against this backdrop, model interconnection enables the sharing, integration, and complementation of large model capabilities by constructing a collaborative paradigm among models. The cascade architecture represents a typical form of such collaboration, where multiple large models are organized in a chain-like sequence to enhance system performance through step-by-step optimization. Routing in model cascades aims to select appropriate cascade paths and serves as a key factor in improving system capabilities. However, current routing evaluation and selection methods lack systematic consideration of model collaboration relationships. To address this, this study proposes a dynamic routing method based on collaboration relationships. It first builds a model collaboration graph through a mutual evaluation mechanism, and then employs a dynamic collaborative routing algorithm to analyze responses hop by hop and optimize path selection. The mutual evaluation mechanism uses gradient-based mutual assessment to quantify the quality of pairwise model collaboration. Based on the resulting collaboration quality information, the dynamic collaborative routing algorithm adopts a model “consensus rule” to analyze each hop’s response and determine the routing order, thus enabling dynamic path adjustment. Experimental results show that the proposed routing algorithm outperforms both non-preset and non-targeted routing methods in terms of accuracy and response win rate on benchmark task datasets. On the OMGEval dataset, the win rate is improved by up to 45% compared to non-preset routing.

Abstract:Deep reinforcement learning has achieved significant breakthroughs in various fields, with policy gradient algorithms widely adopted due to their suitability for handling nonlinear and high-dimensional state spaces. However, in practical applications, existing policy gradient algorithms still suffer from high variance, which slows convergence and may cause suboptimal solutions. To tackle this challenge, a variance optimization method for policy gradients is proposed from a latent variable causal model perspective. By introducing latent variables to characterize unobserved random information, a latent variable causal model is constructed and learned. Utilizing this model, a causal value function is proposed and combined with long short-term memory (LSTM) networks to differentiate the temporal impact of unobserved information on value estimation. This approach improves the accuracy of action advantage function estimation and reduces policy gradient variance. Experiments demonstrate that the proposed latent variable causal model outperforms state-of-the-art algorithms across multiple tasks, with better performance and stability.

Abstract:Recommendation systems have become a key technology in mitigating information overload in the era of big data, with widespread applications in E-commerce and other fields. However, traditional centralized data collection methods expose significant risks of user privacy leakage. Federated learning enables collaborative model training across multiple data holders without the need to share raw user data, thus protecting privacy. Federated recommendation systems have gained considerable attention from both academia and industry. Existing federated recommendation algorithms place the model training process in a distributed environment, effectively avoiding the centralized storage of sensitive user data on a single server. However, these approaches still face challenges related to privacy leakage and high communication costs. To address these issues, this study proposes a communication-efficient federated recommendation algorithm based on differential privacy. The algorithm introduces a general sub-model selection strategy that strengthens privacy protection of user interaction data on the client side through a randomized response mechanism. On the server side, it employs maximum likelihood estimation to infer the true interaction frequencies of items and optimize the sub-model selection process. This strategy achieves an effective balance between privacy protection and model utility. The proposed algorithm is applicable not only to matrix factorization-based recommendation models but also to deep learning-based models, demonstrating high flexibility and adaptability across various recommendation scenarios. Furthermore, to reduce communication overhead, a global model partitioning strategy is proposed to address the complex structures and large parameter sizes of deep learning models. Differentiated optimization strategies are applied to shallow and deep networks to effectively mitigate communication costs. Theoretical analysis shows that the method satisfies differential privacy, while experimental results on real-world datasets demonstrate that the proposed approach preserves user data privacy without significantly compromising model utility, while substantially improving communication efficiency in federated recommendation systems.

Abstract:With the rapid development of the digital economy, data, as an important productive factor, plays a central role in fostering new forms of productive forces and ensuring the efficient functioning of the economy and society. The efficient utilization of data as a production factor, along with its orderly, secure, and compliant circulation among various entities, is essential to realizing its full value. However, data circulation is confronted with significant security risks and trust-related challenges, including leakage, tampering, and unavailability. At the same time, the requirements for data authenticity and circulation transparency are becoming increasingly important. To address these issues, technologies such as cryptography, blockchain, and trusted execution environments are widely applied. However, these technologies still face limitations such as high costs and insufficient flexibility, and current research often focuses on specific stages, lacking a holistic perspective. Therefore, this study is structured around the full life cycle of data factors, develops an analytical framework comprising six stages: data collection, data transmission, data storage, data processing, data publication, and data traceability, and systematically analyzes the security and trust-related challenges at each stage. For the first time, this study explicitly proposes two evolutionary models: the “three-line differentiation and collaborative integration” model for security technologies and the “verification-depth progression” model for trustworthy technologies. This study aims to offer systematic solutions to key issues in this field and outline directions for future research.

Abstract:The computation of signatures is typically performed on physically insecure devices such as mobile phones or small IoT devices, which may lead to private key exposure and subsequently compromise the entire cryptographic system. Key-insulated signature schemes serve as a method to mitigate the damage caused by private key exposure. In a key-insulated cryptosystem, the public key remains constant throughout the entire time period, and the fixed private key is stored on a physically secure device. At the beginning of each time period, the insecure device interacts with the physically secure device storing the fixed private key to obtain the temporary private key for the current time slice. A secure identity-based key-insulated signature scheme must satisfy both unforgeability and key insulation. Key insulation ensures that even if an adversary obtains temporary private keys for multiple time periods, they cannot forge signatures for other periods. SM9 is a commercial identity-based cryptographic standard independently developed by China. This study applies the key-insulated method to the SM9 identity-based signature scheme to resolve the private key exposure issue present in the original scheme. First, a security model for identity-based key-insulated signatures is presented. Then, an identity-based key-insulated signature scheme based on SM9 is constructed. Finally, detailed security proofs and experimental analysis are provided.

Abstract:In privacy-preserving inference using convolutional neural network (CNN) models, previous research has employed methods such as homomorphic encryption and secure multi-party computation to protect client data privacy. However, these methods typically suffer from excessive prediction time overhead. To address this issue, an efficient privacy-preserving CNN prediction scheme is proposed. This scheme exploits the different computational characteristics of the linear and non-linear layers in CNNs and designs a matrix decomposition computation protocol and a parameterized quadratic polynomial approximation for the ReLU activation function. This enables efficient and secure computation of both the linear and non-linear layers, while mitigating the prediction accuracy loss caused by the approximations. The computations in both the linear and non-linear layers can be performed using lightweight cryptographic primitives, such as secret sharing. Theoretical analysis and experimental results show that, while ensuring security, the proposed scheme improves prediction speed by a factor of 2 to 15, with only about a 2% loss in prediction accuracy.

Abstract:Quality of service (QoS)-aware cloud API recommendation systems play an important role in solving cloud API overload problems, differentiating cloud API performance, and achieving high-quality cloud API selection. However, due to the openness of the network environment and the monetary nature of cloud APIs, recommendation systems are susceptible to poisoning attacks, which causes the recommendation results to deviate from fairness and credibility. Existing defense methods against poisoning attacks mainly adopt the “detection and defense” strategy, which utilizes detection algorithms to filter out malicious users before model training to mitigate the influence of the attacks. However, due to the performance limitations of detection algorithms, it is inevitable that malicious users cannot be completely filtered out. To this end, this study proposes a continuous defense method against poisoning attacks on the QoS-aware cloud API recommendation system from a “learning to defense by attacks” perspective with trusted data augmentation. First, this study establishes a defense framework against poisoning attacks based on trusted data augmentation and enhances the robustness of the recommendation system by generating high-quality trusted user data for model training. Second, the study designs a trusted user generation algorithm based on the diffusion model, which employs iterative denoising to learn real-world QoS data distribution related to cloud APIs and generate high-quality trusted user vectors, thus mitigating the influence of data subjected to poisoning attacks on training models. Finally, extensive experiments are conducted based on real-world cloud API QoS datasets, and 11 recommendation algorithms from three categories are utilized to comprehensively evaluate the effectiveness and universality of the proposed defense method. Experimental results indicate that the proposed framework of continuous defense against poisoning attacks based on trusted data augmentation is effective, and the generated trusted user can significantly improve the robustness of the cloud API recommendation system.