Abstract:In certain designs and applications of practical lattice-based cryptography, the use of a specialized variant of LWE problems, where the public matrix is sampled from a non-uniform distribution, is required to establish the securities of corresponding cryptographic schemes. Recently, the formal definition of LWE problems with semi-uniform seeds was introduced in some work, in which the hardness of Euclidean, ideal, and module lattice-based LWE problems with semi-uniform seeds was proved through reduction roadmaps similar to those employed in the hardness proofs of entropic LWE problems. However, known reduction introduces significant losses in the Gaussian parameters of errors and dimensions. Moreover, additional non-standard assumptions are required to demonstrate the hardness of LWE problems with semi-uniform seeds over rings. In this study, a tighter reduction is proposed for LWE problems with semi-uniform seeds by incorporating modified techniques from the hardness proofs of Hint-LWE problems. The proposed reduction is largely unaffected by the algebraic structures of the underlying problems and can be uniformly applied to Euclidean, ideal, and module lattice-based LWE problems with semi-uniform seeds. The hardness of these LWE problems can be established based on standard LWE assumptions without the need for any additional non-standard assumptions. Furthermore, the dimension of the corresponding LWE problems remains unchanged, and the reduction introduces only minimal losses in Gaussian parameters of errors.

Abstract:SPHINCS⁺ is a stateless digital signature scheme designed using hash functions and has been proven resistant to quantum computing attacks. However, its wide practical application is constrained by the large size of the generated signature values. To address the issue of the lengthy signature value generated by the WOTS+ one-time signature scheme within SPHINCS⁺, a compact one-time signature scheme, SM3-OTS, based on Chinese cryptographic algorithm SM3, is proposed in this study. The proposed scheme utilizes the binary and hexadecimal information of the message digest as the indices for node positions in the first 32 hash chains and the last 16 hash chains, respectively. This approach effectively reduces the key length and the signature value length compared to traditional one-time signature schemes based on hash functions. Compared to WOTS+ in SPHINCS⁺, Balanced WOTS+ in SPHINCS-α, and WOTS+C in SPHINCS⁺C, the proposed SM3-OTS shortens the signature value length by about 29%, 27%, and 26%, respectively, with a significant improvement in signing performance. In addition, by adopting the SM3 algorithm, SM3-OTS exhibits strong resistance to quantum attacks while maintaining well-balanced overall performance.

Abstract:Kyber, a key encapsulation mechanism based on lattice problems, was the first to be standardized by the National Institute of Standards and Technology (NIST) in 2023. Kyber-AKE, a weak forward-secure authenticated key exchange (AKE) protocol, was constructed by the designers of Kyber and derives session keys in two rounds using three IND-CCA secure key encapsulation mechanisms. This study introduces Kyber-PFS-AKE, a newly proposed authenticated key exchange protocol. In Kyber-PFS-AKE, only IND-CPA secure public-key encryption is utilized, and decryption errors within IND-CPA secure encryption are addressed using the re-encryption technique within the FO transformation, thus simplifying the design of post-quantum Kyber-AKE. A rigorous proof demonstrates that certain operations in the Kyber-AKE protocol are redundant. By eliminating these redundancies, the protocol achieves a simpler and more efficient design. The session key indistinguishability and perfect forward security of Kyber-PFS-AKE are formally proven within the eCK-PFS-PSK model. The proposed Kyber-PFS-AKE is implemented using Kyber-768. PKE with 165-bit quantum security. Experimental results show that compared to Kyber-AKE, the computation time for the initiator is reduced by 38%, while the computation time for the responder is reduced by 30%.

Abstract:Recognized as a crucial privacy-protecting technology, group signatures provide robust anonymity assurances for users. However, conventional group signature schemes often rely on group managers capable of revealing the identities of signers, a feature that contradicts the decentralized nature of blockchain and fails to meet stricter privacy demands in certain applications. To address these limitations, this study introduces a group signature scheme with user-controlled linkability and verifier conditional revocation, inspired by double-authentication-preventing signatures and existing linkable and revocable group signatures. The proposed scheme achieves an optimal balance between user privacy and platform oversight, with a concrete instantiation constructed on lattices. Under the random oracle model, the scheme is demonstrated to satisfy the properties of selfless anonymity, traceability, and non-frameability. Performance evaluations indicate that both time and communication costs remain within acceptable limits, ensuring feasibility for practical deployment. In addition, a post-quantum secure medical data sharing system has been designed, integrating the proposed group signature scheme with blockchain technology.

Abstract:Digital signature algorithms play a vital role in network security infrastructure. The majority of current digital signature schemes rely on RSA and ECC. However, with the rapid advancement of quantum computing, traditional public-key cryptographic schemes face increasing security risks. As a result, researching and deploying cryptographic schemes capable of resisting quantum attacks has become a critical research direction. Following multiple rounds of evaluation and analysis, National Institute of Standards and Technology (NIST) announced the post-quantum digital signature standard ML-DSA in August 2024, with Dilithium as its core algorithm. In light of the high-dimensional polynomial matrix operations characteristic of Dilithium, this study proposes various optimization strategies based on the FPGA platform. These include multifunctional systolic array operation units with configurable parameters, dedicated polynomial parallel sampling modules, reconfigurable storage units designed for multiple parameter sets, and high-parallelism timing state machines tailored for complex multi-module architectures. These optimizations aim to overcome performance bottlenecks and achieve enhanced signature operation efficiency, ultimately realizing a digital signature hardware architecture that supports three security levels simultaneously. The proposed hardware architecture is deployed and evaluated on the Xilinx Artix-7 FPGA platform and compared against existing implementations. The results demonstrate that the proposed design achieves improvements in signature operation efficiency by factors of 7.4, 8.3, and 5.6 across the three security levels, respectively. This advancement provides a robust performance foundation for quantum-resistant digital signature applications and offers valuable insights for the engineering and practical deployment of lattice cryptographic schemes.

Abstract:With the widespread application of blockchain technology, authenticated storage, as a core component, plays a crucial role in ensuring data integrity and consistency. In traditional blockchain systems, authenticated storage is maintained through a series of cryptographic algorithms, which verify transactions and preserve the integrity of ledger states. However, the advent of quantum computers has introduced a significant threat to existing blockchain authentication storage technologies, raising the risk of data breaches and compromised integrity. The most advanced authenticated storage schemes primarily rely on the bilinear Diffie-Hellman assumption, which is susceptible to quantum attacks. To enhance the security and efficiency of authenticated storage, this study introduces a stateless hash signature mechanism and proposes the quantum-resistant blockchain authenticated storage scheme EQAS. The proposed scheme decouples data storage from data authentication, utilizes random forest chains to efficiently generate commitment proofs, and employs a hyper tree structure to perform efficient authentication. Security analyses show that EQAS is resistant to quantum algorithm attacks. Comparative experiments with other authenticated storage schemes demonstrate the superior efficiency and performance of EQAS in handling blockchain authentication storage tasks.

Abstract:With the development of quantum computers, public blockchains relying on traditional elliptic curve digital signatures are expected to face disruptive security risks. A common solution involves the application of post-quantum digital signature algorithms within blockchain systems. For public blockchains utilizing the proof-of-work consensus mechanism, ensuring sufficient computing power is regarded as a critical foundation for security. Energy conservation and the maximization of computing power support have been identified as key research directions. Therefore, a post-quantum blockchain system featuring diversified computing power and autonomous post-quantum signature is proposed in this study. The Dilithium signature scheme, recommended by the National Institute of Standards and Technology (NIST) as a preferred and general-purpose post-quantum signature standard, relies on the security of MLWE and MSIS problems in power-of-two cyclotomic rings. However, similar to the early adoption of the EC-DSA standard in Bitcoin without adherence to the NIST-specific elliptic curves, the rich algebraic structure of power-of-two cyclotomic rings poses greater risks and uncertainties regarding long-term security. To address this, a more conservative and secure approach, based on post-quantum lattice-based cryptography with fewer algebraic structures, is constructed. In this study, a Dilithium variant, Dilithium-Prime, based on a large-Galois-group prime-degree prime-ideal field, is proposed as the signature algorithm for the post-quantum blockchain system to ensure high-confidence transaction signing with post-quantum security. To maximize the computing power support for the post-quantum public blockchain and address the current issue of declining mining pool and miner income, a multi-parent chain auxiliary proof-of-work consensus mechanism is introduced. This mechanism enables the request of computing power from all miners using Sha256 and Scrypt hash calculations to assist in consensus without increasing the workload for existing miners and mining pools. As a result, the source of computing power for the post-quantum blockchain is expanded, and the utilization rate of existing mining pools and miners is improved. In addition, a block and transaction structure, along with a difficulty adjustment algorithm tailored for this multi-parent chain auxiliary proof-of-work consensus mechanism, is proposed. This system stabilizes the block production ratio and production time across different levels of computing power and effectively responds to extreme cases, such as sudden surges or reductions in computing power, ensuring the system’s robustness.

Abstract:This study discusses the computational complexity of the partition function of the symmetric dual-spin system on regular graphs. Based on # exponential time hypothesis (#ETH) and random exponential time hypothesis (rETH), this study develops the classical dichotomies of this problem class into the exponential dichotomies, also known as the fine-grained dichotomies. In other words, this study proves that when the given tractable conditions are satisfied, then the problem is solvable in polynomial time; otherwise, there is no sub-exponential time algorithm when #ETH holds. This study also proposes two solutions to solve the in-effectiveness of existing interpolation methods on building sqrt-sub-exponential time reductions under the restriction of planar graphs. It also utilizes these two solutions to discuss the related fine-grained complexity and dichotomy of this problem under the planar graph restriction.

Abstract:Scalar multiplication is the core operation in traditional elliptic curve cryptography (ECC). Scalar representations determine the iterations in scalar multiplication algorithms, which directly affect the security and efficiency of the algorithms. This study proposes two new scalar representation algorithms. One algorithm is ordered window width non-adjacent form (OWNAF) which combines traditional window non-adjacent form with random key segmentation and can resist energy analysis attacks while yielding better efficiency. The other is called window joint regular form (wJRF), which is improved from the traditional joint regular form. The wJRF algorithm is applicable to multi-scalar multiplication algorithms, which can reduce computational costs and ensure sound security compared with the existing algorithms.

Abstract:In recent years, as an algorithm for identifying bug-introducing changes, SZZ has been widely employed in just-in-time software defect prediction. Previous studies show that the SZZ algorithm may mislabel data during data annotation, which could influence the dataset quality and consequently the performance of the defect prediction model. Therefore, researchers have made improvements to the SZZ algorithm and proposed multiple variants of SZZ. However, there is no empirical study to explore the effect of data annotation quality by SZZ on the performance and interpretability of just-in-time defect prediction for mobile APP. To investigate the influence of mislabeled changes by SZZ on just-in-time defect prediction for mobile APP, this study conducts an extensive and in-depth empirical comparison of four SZZ algorithms. Firstly, 17 large-scale mobile APP projects are selected from the GitHub repository, and software metrics are extracted by adopting the PyDriller tool. Then, B-SZZ (original SZZ), AG-SZZ, MA-SZZ, and RA-SZZ are employed for data annotation. Then, the just-in-time defect prediction models are built with random forest, naive Bayes, and logistic regression classifiers based on the time-series data partitioning. Finally, the performance of the models is evaluated by traditional measures of AUC, MCC, and G-mean, and effort-aware measures of F-measure@20% and IFA, and a statistical significance test and interpretability analysis are conducted on the results by employing SKESD and SHAP respectively. By comparing the annotation performance of the four SZZ algorithms, the results are as follows. (1) The data annotation quality conforms to the progressive relationship among SZZ variants. (2) The mislabeled changes by B-SZZ, AG-SZZ, and MA-SZZ can cause performance reduction of AUC and MCC of different levels, but cannot lead to performance reduction of G-mean. (3) B-SZZ is likely to cause a performance reduction of F-measure@20%, while B-SZZ, AG-SZZ, and MA-SZZ are unlikely to increase effort during code inspection. (4) In terms of model interpretation, different SZZ algorithms will influence the three metrics with the largest contribution during the prediction, and the la metric has a significant influence on the prediction results.

Abstract:With the continuous development of information technology, the quantity and variety of software products are increasing, but even high-quality software may contain vulnerabilities. In addition, the software update speed is fast, and the software architecture is increasingly complex, which leads to the gradual evolution of vulnerabilities into new forms. Consequently, traditional vulnerability detection methods and rules are difficult to apply to new vulnerability features. Due to the scarcity of zero-day vulnerability samples, zero-day vulnerabilities that appear in the software evolution process are difficult to find, which brings great potential risks to software security. This study proposes a vulnerability sample generation method based on abstract syntax tree mutation, which can simulate the structure and syntax rules of real vulnerabilities, generate vulnerability samples more in line with the actual situation, and provide a more effective solution for software security and reliability. This method analyzes the abstract syntax tree structure generated by Eclipse CDT, extracts the syntactic information in the nodes, reconstructs the nodes and abstract syntax trees, optimizes the abstract syntax tree structure, and designs a series of mutation operators. Subsequently, it performs mutation operations on the optimized abstract syntax trees. The method proposed in this paper can generate mutation samples with the characteristics of UAF and CUAF vulnerabilities, which can be used for the detection of zero-day vulnerabilities and help to improve the detection rate of zero-day vulnerabilities. Experimental results show that this method reduces the invalid sample size by 34% on average compared with the random variation method in traditional detection methods, and can generate more complex mutated samples. In addition, this method can generate more complex mutated samples, enhancing the coverage and accuracy of detection.

Abstract:To address the issue of untrustworthy behaviors resulting from malicious attackers exploiting security vulnerabilities within smart contracts in the consortium blockchain system, this study introduces a trusted verification mechanism of smart contract behavior for consortium blockchain to conduct trusted verification for contract behavior integrity. Firstly, the proposed approach takes the system call as the smallest behavior unit and describes the historical behavioral state with the behavior sequence based on system calls. Subsequently, on the premise of ensuring the trustworthiness of contract code release and the execution environment, it performs trusted verification according to predefined behavioral rules during contract execution. Finally, a theoretical analysis of this mechanism is carried out, and an experimental evaluation is conducted in the Hyperledger Fabric environment. Results demonstrate that the proposed method can effectively achieve the trusted verification of smart contract behavior and ensure the credibility of behavior within the life cycle of smart contracts.

Abstract:Blockchain has shown strong vitality in the field of cryptocurrency investment, attracting the participation of a large number of investors. However, due to the anonymity of blockchain, it induces a lot of fraud, among which the Ponzi scheme smart contract is a typical fraudulent investment activity, causing huge economic losses for investors. Therefore, the detection of Ponzi scheme contracts on Ethereum becomes particularly important. Nevertheless, most existing studies have ignored control flow information in the source code of Ponzi scheme contracts. To extract more semantic and structural information from Ponzi scheme contracts, this study proposes a Ponzi scheme contract detection model based on code control flow graph. First, the model constructs the obtained contract source code in the form of a control flow diagram. Then, key features including data flow information and code structure information are extracted by the Word2Vec algorithm. Considering that the functions of each smart contract are different and the length of the code varies significantly, resulting in a large difference in the extracted feature vectors. In this study, feature vectors generated by different smart contracts are aligned so that all feature vectors have the same dimension, which is convenient for subsequent processing. Secondly, the feature learning module based on graph convolution and Transformer is utilized to introduce multi-head attention mechanism to learn the dependency of node features. Finally, the multilayer perceptron is used to identify the Ponzi scheme contract. By comparing the proposed model with the traditional graph feature learning model on the dataset provided by the XBlock website, the performance of the multi-head attention mechanism introduced by the model is verified. Experimental results demonstrate that this model effectively improves the ability to detect Ponzi scheme contracts.

Abstract:The application of artificial intelligence technology has extended from relatively static tasks such as classification, translation, and question answering to relatively dynamic tasks that require a series of “interaction-action” with the environment to be completed, like autonomous driving, robotic control, and games. The core of the model for executing such tasks is the sequential decision-making (SDM) algorithm. As it faces higher uncertainties of the environment and interaction and these tasks are often safety-critical systems, the testing techniques are confronted with great challenges. The existing testing technologies for intelligent algorithm models mainly focus on the reliability of a single model, the generation of diverse test scenarios for complex tasks, simulation testing, etc., while no attention is paid to the “interaction-action” decision sequence of the SDM model, leading to unadaptability or low cost-effectiveness. In this study, a fuzz testing method named IIFuzzing for intervening in the execution of inert “interaction-action” decision sequences is proposed. In the fuzz testing framework, by learning the “interaction-action” decision sequence pattern, the inert “interaction-action” decision sequences that will not trigger failure accidents are predicted and the testing execution of such sequences is terminated to improve the testing efficiency. The experimental evaluations are conducted in four common test configurations, and the results show that compared with the latest fuzz testing for SDM models, IIFuzzing can detect 16.7%–54.5% more failure accidents within the same time, and the diversity of accidents is also better than that of the baseline approach.

Abstract:Deep stochastic configuration network (DSCN) adopts a feedforward learning approach and randomly assigns node parameters based on a unique supervisory mechanism, which has universal approximation. However, in actual scenarios, the potential outliers and noise during data collection can negatively affect the classification results. To improve the performance of DSCN in solving binary classification problems, this study introduces the idea of intuitionistic fuzzy numbers based on DSCN and proposes an intuitionistic fuzzy deep stochastic configuration network (IFDSCN). Different from the standard DSCN, IFDSCN assigns an intuitionistic fuzzy number to each sample by calculating the sample membership and non-membership, and generates the optimal classifier by a weighting method to overcome the negative effect of noise and outliers on data classification. The experimental results on eight benchmark datasets show that compared to other learning models including the intuitionistic fuzzy twin support vector machine (IFTWSVM), kernel ridge regression (KRR), intuitionistic fuzzy kernel ridge regression (IFKRR), random vector functional link neural network (RVFL), and SCN, IFDSCN has better binary classification performance.

Abstract:Business rules are crucial for the securities domain and serve as the source of requirements for securities trading systems. Due to the variability of these business rules, how to improve the efficiency of specifying software requirements from business rule trading documents has become a core problem. The securities business rule documents feature numerous software-unrelated descriptions, abundant professional terms, and many context-related expressions and abstract representations, which necessitate the support of domain-specific knowledge for automatic specification. As a result, how to integrate the domain-related knowledge into the automatic process becomes a key problem for specification. This study proposes an automatic specification method for securities domain businesses integrating large language models and the domain knowledge base. It leverages the large language models, employing techniques such as fine-tuning and in-context learning to embed domain knowledge for natural language processing tasks such as rule classification and requirement information extraction. Additionally, this study also employs the domain knowledge base to provide professional knowledge and assist in the operationalization and relationship extraction of requirements. Finally, requirement specification in the form of data flow is formed. The evaluation results show that the proposed approach can process business rule documents in various securities trading fields, achieving an average function point identification rate of 91.97% on the evaluation dataset, which matches or even surpasses the level of experts in the domain, with the efficiency improved by an average of 10 times compared to human participants.

Abstract:Offline reinforcement learning has yielded significant results in tasks with continuous and intensive rewards. However, since the training process does not interact with the environment, the generalization ability is reduced, and the performance is difficult to guarantee in a discrete and sparse reward environment. The diffusion model combines the information in the neighborhood of the sample data with noise addition to generate actions that are close to the distribution of the sample data, which strengthens the learning and generalization ability of the agents. To this end, offline reinforcement learning with diffusion models and expectation maximization (DMEM) is proposed. The method updates the objective function by maximizing the expectation of the maximum likelihood logarithm to make the strategy more generalizable. Additionally, the diffusion model is introduced into the strategy network to utilize the diffusion characteristics to enhance the ability of the strategy to learn data samples. Meanwhile, the expectile regression is employed to update the value function from the perspective of high-dimensional space, and a penalty term is introduced to make the evaluation of the value function more accurate. DMEM is applied to a series of tasks with discrete and sparse rewards, and experiments show that DMEM has a large advantage in performance over other classical offline reinforcement learning methods.

Abstract:Large language model (LLM) like ChatGPT has found widespread applications across various fields due to their strong natural language understanding and generation capabilities. However, deep learning models exhibit vulnerability when subjected to adversarial example attacks. In natural language processing, current research on adversarial example generation methods typically employs CNN-based models, RNN-based models, and Transformer-based pre-trained models as target models, with few studies exploring the robustness of LLMs under adversarial attacks and quantifying the evaluation criteria of LLM robustness. Taking ChatGPT against Chinese adversarial attacks as an example, this study introduces a novel concept termed offset average difference (OAD) and proposes a quantifiable LLM robustness evaluation metric based on OAD, named OAD-based robustness score (ORS). In a black-box attack scenario, this study selects nine mainstream Chinese adversarial attack methods based on word importance to generate adversarial texts, which are then employed to attack ChatGPT and yield the attack success rate of each method. The proposed ORS assigns a robustness score to LLMs for each attack method based on the attack success rate. In addition to the ChatGPT that outputs hard labels, this study designs ORS for target models with soft-labeled outputs based on the attack success rate and the proportion of misclassified adversarial texts with high confidence. Meanwhile, this study extends the scoring formula to the fluency assessment of adversarial texts, proposing an OAD-based adversarial text fluency scoring method, named OAD-based fluency score (OFS). Compared to traditional methods requiring human involvement, the proposed OFS greatly reduces evaluation costs. Experiments conducted on real-world Chinese news and sentiment classification datasets to some extent initially demonstrate that, for text classification tasks, the robustness score of ChatGPT against adversarial attacks is nearly 20% higher than that of Chinese BERT. However, the powerful ChatGPT still produces erroneous predictions under adversarial attacks, with the highest attack success rate exceeding 40%.

Abstract:Twin support vector machine (TSVM) can effectively tackle data such as cross or XOR data. However, when set-valued data are handled, TSVM usually makes use of statistical information of set-valued objects such as the mean and the median. Unlike TSVM, this study proposes twin support function machine (TSFM) that can directly deal with set-valued data. In terms of support functions defined for set-valued objects, TSFM obtains nonparallel hyperplanes in a Banach space. To suppress outliers in set-valued data, TSFM adopts the pinball loss function and introduce the weights of set-valued objects. Considering that TSFM involves optimization problems in the infinite-dimensional space, the measure is taken in the form of a linear combination of Dirac measures. Thus the optimization model in the finite-dimensional space is constructed. To solve the optimization model effectively, this study employs the sampling strategy to transform the model into quadratic programming (QP) problems. The dual formulations of the QP problems are derived, which provides theoretical foundations for determining which sampling points are support vectors. To classify set-valued data, the distance from the set-valued object to the hyperplane in a Banach space is defined, and the decision rule is derived therefrom. This study also considers the kernelization of support functions to capture the nonlinear features of data, which makes the proposed model available for indefinite kernels. Experimental results demonstrate that TSFM can capture the intrinsic structure of cross-plane set-valued data and obtain good classification performance in the case of outliers or set-valued objects containing a few high-dimensional examples.

Abstract:Recently, aspect sentiment understanding in conversational texts has received growing attention from researchers, leading to notable progress in the field. Unlike prior research, this study focuses on the performance of large language models (LLMs) in conversational aspect sentiment understanding tasks and highlights two primary challenges: aspect-coreference mapping and aspect-sentiment mapping. These challenges pose significant constraints on achieving accurate aspect sentiment understanding within conversational structures. To address these issues, the study defines a new task for LLMs in conversational aspect sentiment understanding, aiming to extract quadruples that encapsulate both aspect-coreference mapping and aspect-sentiment mapping relationships. A high-quality annotated dataset of conversational aspect sentiment quadruples has been created to evaluate the performance of LLMs on this task. To tackle the challenges of mapping relationships and mitigate the inherent hallucination issues of LLMs, this study introduces a multi-agent consistency reflection approach. This method involves the design of three sub-task agents to aid in capturing the complex mapping relationships within conversational structures. In addition, an enhanced consistency reflection is proposed, enabling the model to generate optimal results through multi-agent consensus, thus alleviating hallucination problems. Experimental results demonstrate that the proposed approach significantly outperforms state-of-the-art benchmarks. Furthermore, it exhibits superior capabilities in extracting aspect referential relationships and aspect sentiments, contributing to advancements in fine-grained sentiment understanding within conversational contexts using LLMs.

Abstract:Brain-inspired computing chips of various architectures are emerging, and the inference/training/learning algorithms of spiking neural network (SNN) and the efficient simulation of biological neural networks have become research hotspots. Meanwhile, efficiently executing applications with different computation/memory-access characteristics on various chips remains a significant challenge, which is crucial for establishing a robust brain-inspired computing ecosystem. The success of the general-purpose computing ecosystem indicates that a flexible, scalable, and reusable compiler infrastructure is an effective solution to this problem. This study proposes BIVM, a compilation framework for brain-inspired computing, along with its proof-of-concept implementation. Based on the multi-level intermediate representation (MLIR) framework of domain specific architecture (DSA), multi-layer IRs customized for SNNs are designed, including an SNN dialect, middle-layer IRs composed mainly of MLIR’s inherent dialects, and the underlying IRs for various target chips. To address challenges such as the large architectural differences and varying granularity of hardware primitives in brain-inspired chips, BIVM leverages MLIR’s progressivity feature. This allows for the mixing of different abstraction levels and concepts (e.g. combining fine-grained instructions with coarse-grained computation based on the crossbar structure specific to certain back-ends), enabling software module reuse and reducing compiler development costs, ultimately leading to high productivity. In addition, the framework provides flexibility to combine various levels of compilation optimizations, including widely-used SNN-specific optimizations (e.g. exploring computing sparsity and improving parallelism) and low-level optimizations tailored to different back-ends, ensuring performance portability. The current BIVM prototype supports back-ends such as general-purpose processors (control-flow architecture), SNN accelerator chips (FPGAs) with a hybrid control-/data-flow architecture, and data-flow chip designs based on ReRAM (resistive random-access memory, a widely-used neuromorphic device). It can optimize and compile deep SNN and biological neural network simulation applications into executables tailored for these chips. Comprehensive testing and performance comparisons demonstrate the potential of this compilation framework in achieving high productivity, portability, and performance.

Abstract:The resource auction mechanism can maximize the resource allocation benefit by fully introducing competition, and has found widespread applications in mobile edge computing resource allocation and pricing. Currently, auction mechanisms relevant to resource allocation of mobile edge computing mainly focus on computing resource allocation of edge servers, and there are limitations in both considering the allocation of wireless bandwidth resources that do not belong to any edge servers and computing resources belonging to specific edge servers under a multi-base station environment. Furthermore, with multiple types of resource conditions considered, a challenge is posed to the design of a resource allocation and pricing strategy that guarantees benefits for both resource providers and users. By analyzing the characteristics of multi-base stations and resource constraints, this study proposes a double-auction-based combinational resource allocation (DACRA) mechanism for mobile edge computing. This mechanism considers the allocation of wireless bandwidth resources in multi-communication base stations and multiple computing resources of edge servers and introduces resource scarcity and bidding density to ensure high allocation efficiency. Theoretical analysis shows that the DACRA mechanism is a polynomial time algorithm that satisfies incentive compatibility, budget-balance, and individual rationality. Simulation results based on a publicly available dataset show that the proposed mechanism can yield lower computational time costs, and higher social welfare, request success rates, and resource utilization rates than existing research results.

Abstract:With the continuous deepening of research on the security and privacy of deep learning models, researchers find that model stealing attacks pose a tremendous threat to neural networks. A typical data-dependent model stealing attack can use a certain percentage of real data to query the target model and train an alternative model locally to steal the target model. Since 2020, a novel data-free model stealing attack method has been proposed, which can steal and attack deep neural networks simply by using fake query examples generated by generative models. Since it does not rely on real data, the data-free model stealing attack can cause more serious damage. However, the diversity and effectiveness of the query examples constructed by the current data-free model stealing attack methods are insufficient, and there are problems of a large number of queries and a relatively low success rate of the attack during the model stealing process. Therefore, this study proposes a vision feature decoupling-based model stealing attack (VFDA), which decouples and generates the visual features of the query examples generated during the data-free model stealing process by using a multi-decoder structure, thus improving the diversity of query examples and the effectiveness of model stealing. Specifically, VFDA uses three decoders to respectively generate the texture information, region encoding, and smoothing information of query examples to complete the decoupling of visual features of query examples. Secondly, to make the generated query examples more consistent with the visual features of real examples, the sparsity of the texture information is limited and the generated smoothing information is filtered. VFDA exploits the property that the representational tendency of neural networks depends on the image texture features, and can generate query examples with inter-class diversity, thus effectively improving the similarity of model stealing and the success rate of the attack. In addition, VFDA adds intra-class diversity loss to the smoothed information of query samples generated through decoupling to make the query samples more consistent with real sample distribution. By comparing with multiple model stealing attack methods, the VFDA method proposed in this study has better performance in the similarity of model stealing and the success rate of the attack. In particular, on the GTSRB and Tiny-ImageNet datasets with high resolution, the attack success rate is respectively improved by 3.86% and 4.15% on average compared with the currently better EBFA method.

Abstract:The uBlock algorithm has been widely used in algorithm design, side channel protection, Internet of Things applications, and cryptanalysis. Although the uBlock algorithm is suitable for high-speed implementation, the publicly available implementation rate of this algorithm is far lower than that of algorithms such as AES and SM4. Bit slicing is a common method to optimize block ciphers. However, when using bit slicing to optimize the uBlock algorithm, it faces the problem of huge memory access overhead due to insufficient register resources. In this study, a flexible bit slicing optimization method named FBS-uBlock is designed for the uBlock algorithm. It reduces the number of registers occupied by the algorithm under bit slicing, thus reducing the memory access overhead and improving the speed. After testing, the proposed optimization method can reduce the memory access instruction of uBlock-128/128, uBlock-128/256, and uBlock-256/256 algorithms by up to 71%, 71%, and 72%, respectively. The maximum encryption rates can reach 12758 Mb/s, 8944 Mb/s, and 8984 Mb/s respectively, which are 3.9, 4.2, and 3.4 times higher than the implementation rates in the design documentation.

Abstract:The purpose of text-image person re-identification is to employ the text description to retrieve the target persons in the image database. The main challenge of this technology is to embed image and text features into common potential space to achieve cross-modal alignment. Many existing studies try to adopt separate pre-trained unimodal models to extract visual and text features, and then employ segmentation or attention mechanisms to obtain explicit cross-modal alignment. However, these explicit alignment methods generally lack the underlying alignment ability needed to effectively match multimodal features, and the utilization of preset cross-modal correspondence to achieve explicit alignment may result in modal information distortion. An implicit multi-scale alignment and interaction for text-image person re-identification method is proposed. Firstly, the semantic consistent feature pyramid network is employed to extract multi-scale features of the images, and attention weights are adopted to fuse different scale features including global and local information. Secondly, the association between image and text is learned using a multivariate interaction attention mechanism, which can effectively capture the corresponding relationship between different visual features and text information, narrow the gap between modes, and achieve implicit multi-scale semantic alignment. Additionally, the foreground enhancement discriminator is adopted to enhance the target person and extract purer person features, which is helpful for alleviating the information inequality between images and texts. Experimental results on three mainstream text-image person re-identification datasets of CUHK-PEDES, ICFG-PEDES and RSTPReid show that the proposed method effectively improves the cross-modal retrieval performance, which is 2%-9% higher than the Rank-1 of SOTA algorithm.

Abstract:Existing adversarial example detection methods based on image transformation employ the characteristic that the image transformation can significantly change the feature distribution of adversarial examples but slightly change the feature distribution of benign examples. Adversarial examples can be detected by calculating the feature distance before and after image transformation. However, with the deepening research on adversarial attacks, researchers pay more attention to enhancing the robustness of adversarial examples, so that some attacks can be “immune” to the effect exerted by image transformation. Existing methods are difficult to detect robust adversarial examples effectively. This paper observes that the existing adversarial examples are too robust, and the feature distribution distance of robust adversarial examples under image transformation is much smaller than that of benign examples, which is not consistent with the feature distribution laws of benign examples. Based on this key observation, this study proposes a dual-threshold adversarial example detection based on image transformation, which sets a lower threshold combining existing single-threshold methods to form a dual-threshold detection interval. An example whose feature distribution is not within the dual-threshold detection interval will be judged as an adversarial example. Additionally, this study conducts extensive experiments on VGG19, DenseNet, and ConvNeXt models for image classification. The results show that the proposed approach is compatible with the detection ability of existing single-threshold detection schemes, and yields outstanding detection performance against robust adversarial examples.

Abstract:Binary neural networks (BNNs) are highly appealing to the industry due to their significantly reduced computation and storage requirements. However, their accuracy still lags behind that of networks with full-precision parameters. Most existing methods focus on improving the performance of BNNs through advanced training techniques. Empirical findings reveal that the representation capability of quantized features is considerably weaker than that of full-precision features. To address this limitation, a widening and squeezing mechanism is proposed to construct high-accuracy yet compact BNNs. Specifically, features from the original full-precision networks are projected into high-dimensional quantized features to mitigate the representation gap. Meanwhile, redundant quantized features are pruned to prevent the over growth of feature dimensions. As a result, a compact yet sufficiently expressive quantized neural network is constructed. Experimental results on benchmark datasets demonstrate that the proposed method achieves high-accuracy BNNs with significantly fewer parameters and computations while delivering performance comparable to full-precision baseline models. For instance, the binary ResNet-18 achieves a top-1 accuracy of 70% on the ImageNet dataset.