Sorting is one of the most basic algorithms, and developing highly parallel sorting programs is becoming increasingly important in high-performance computing because the number of CPU cores per node in modern supercomputers tends to increase. In this study, we have implemented two multi-threaded sorting algorithms based on samplesort and compared their performance on the supercomputer Fugaku. The first algorithm divides an input sequence into multiple blocks, sorts each block, and then selects pivots by sampling from each block at regular intervals. Each block is then partitioned using the pivots, and partitions in different blocks are merged into a single sorted sequence. The second algorithm differs from the first one in only selecting pivots, where the binary search is used to select pivots such that the number of elements in each partition is equal. We compare the performance of the two algorithms with different sequential sorting and multiway merging algorithms. We demonstrate that the second algorithm with BlockQuicksort (a quicksort accelerated by reducing conditional branches) for sequential sorting and the selection tree for merging shows consistently high speed and high parallel efficiency for various input data types and data sizes.

Partial redundancy elimination (PRE) is a code optimization algorithm that simultaneously performs common sub-expression elimination and loop-invariant code motion. Traditional PREs analyze the entire program and eliminate redundancies. By contrast, demand-driven PRE (DDPRRE) is proposed as an algorithm to analyze only a part of the program to determine whether each expression is redundant by query propagation. The previous DDPRE reduces the analysis time because it limits the analysis range; however, it is known that redundancy may not be eliminated when the nodes in the loop are revisited. We propose a novel DDPRE, named lazy demand-driven PRE (LDPRE), which eliminates redundancy by delaying the decision of whether the analyzing expression is redundant or not when a node in a loop is revisited and redundancy cannot be analyzed. LDPRE uses a semi-lattice as the answer space during the query propagation. The semi-lattice includes not only true/false implying that the queried expression is redundant or not, but also ⊤ for undecidability. While maintaining the analytical efficiency characteristic of demand-driven analysis, our algorithm eliminates redundancy that previous DDPRE could not eliminate by determining the answer using semi-lattice.

Long Method is a code smell that frequently happens in software development, which refers to the complex method with multiple functions. Detecting and refactoring such problems has been a popular topic in software refactoring, and many detection approaches have been proposed. In past years, the approaches based on metrics or rules have been the leading way in long method detection. However, the approach based on deep learning has also attracted extensive attention in recent studies. In this paper, we propose a graph-based deep learning approach to detect Long Method. The key point of our approach is that we extended the PDG (Program Dependency Graph) into a Directed-Heterogeneous Graph as the input graph and used the GCN (Graph Convolutional Network) to build a graph neural network for Long Method detection. Moreover, to get substantial data samples for the deep learning task, we propose a novel semi-automatic approach to generate a large number of data samples. Finally, to prove the validity of our approach, we compared our approach with the existing approaches based on five groups of datasets manually reviewed. The evaluation result shows that our approach achieved a good performance in Long Method detection.

This paper analyzes the impact of input sparsity and DFS/DVFS configurations for single-board computers on the execution time, power, and energy of each VGG16 layer as the first step towards efficient CNN inference on single-board computers. For this purpose, we first develop a power and execution time measurement environment and perform experiments using Raspberry Pi 4 and NVIDIA Jetson Nano. Our results show that clock frequency strongly correlates with execution time and power. Inversely, input sparsity has a weak correlation with execution time and power. Then, we show that a coarse-grained DVFS model can explain over 96% of the variations in the power of each VGG16 layer even when sets of clock frequency and voltage on the single-board computer are unavailable.

Self-timed pipelines (STPs) are becoming attractive because of their power performance efficiency. A circular STP which realizes a looped data flow is necessary to directly implement not only iterative or recursive operations but also circular data paths for program execution. To facilitate product development or prototyping of STP circuits on a commercial field-programmable gate array (FPGA), several research efforts have already made it possible to utilize industry-standard electronic design automation (EDA) tools. However, how to adequately achieve a circular STP whose data transfer is realized by a so-called four-phase bundled-data is still unknown. In this paper, we point out that conventional circuits lead to a design failure or even unacceptably deteriorated throughput because EDA tools improperly interpret their configuration, especially in the realization of functions such as pipeline branching and a data copy and erasure. We propose a circular STP design method composed of both a low-latency handshake circuit configuration and its design procedure. Our proposed method guides the EDA tools to exploit FPGA's intrinsic low-latency paths. We evaluate a circular STP implementing a data-driven processor under corner conditions and show that our method can extract the maximum throughput of target pipelined circuits, which indicates the circular STPs wider applicability.

DX is a must for Japanese companies to cope with Japan's 2025 Digital Cliff. To promote DX initiatives, self-assessment is necessary to make their position recognized. However, conventional manual assessments have high costs, hindering their DX promotion. We noticed that progress in corporate DX can be reflected in the amount of related information on the Web, and there is no reported method of utilizing this relation for DX assessment. This paper proposes WISDOM-DX, a system that accumulates information on corporate activities on the Web using a question-answering system, and ranks companies regarding DX initiatives. We compared WISDOM-DX with two baselines Prize and Google. The precision of WISDOM-DX, Prize, and Google were 56.3%, 45.8%, and 22.9%, respectively. The rate of DX-related award winners or certified companies obtained by WISDOM-DX and Google were 91.7% and 64.6%, respectively. The Area Under the Precision-Recall curve (AUPR) of WISDOM-DX, Prize, and Google were 0.540, 0.359, and 0.181, respectively. An opinion survey showed 60.7% positive and 32.1% neutral responses regarding the agreeability of WISDOM-DX's rankings, and 46.4% positive and 39.3% neutral responses regarding the usefulness of WISDOM-DX. These results showed WISDOM-DX's promising performance and the prospect of automating large-scale assessment regarding corporate DX initiatives.

This paper summarizes our cryptanalysis results on real-world End-to-End Encryption (E2EE) schemes published in recent years. Our targets are LINE (a major messaging application), SFrame (an E2EE protocol adopted by major video/audio applications), and Zoom (a major video communication application). For LINE, we show several attacks against the message integrity of Letter Sealing, the E2EE protocol of LINE, that allow forgery and impersonation. For SFrame, we reveal a critical issue that leads to an impersonation (forgery) attack by a malicious group member with a practical complexity. For Zoom, we discover several attacks more powerful than those expected by Zoom according to their whitepaper. Specifically, if insiders collude with meeting participants, they can impersonate any Zoom user in target meetings, whereas Zoom indicates that they can impersonate only the current meeting participants. We also describe several important works in the area of E2EE security research.

Low-Rate denial of service (LDoS) attacks degrade the quality of service with less traffic than ordinary DoS attacks. LDoS attacks can easily evade conventional counter-DoS detection mechanisms because their time-averaged flow is small and, therefore, become a serious problem. With the recent spread of IoT devices, Zigbee has attracted much attention. Zigbee is a low-power wireless communication protocol that sacrifices transfer range and bandwidth. Since Zigbee consumes very low power, it is widely adopted for small inexpensive IoT devices. The advantage of the low power consumption of Zigbee is due to the indirect transmission. We have already pointed out LDoS attack methods exploiting the characteristics of the indirect transmission, proposed algorithms detecting attackers, and evaluated the accuracy of the algorithms. In this paper, we focus on memory efficient implementation of the algorithm. First, we found that straightforward implementation of the algorithm needs large memory. Then, we propose an improved implementation which requires much less memory. Furthermore, we implement it on a resource-constrained single-board computer and confirm that our proposed algorithm can work correctly with much less memory space and shorter execution time than our previously proposed method. These results prove that the proposed detection algorithm is feasible for a wider range of IoT devices.

Cryptanalysis in a blackbox setting using deep learning is powerful because it does not require the attacker to have knowledge about the internal structure of the cryptographic algorithm. Thus, it is necessary to design a symmetric key cipher that is secure against cryptanalysis using deep learning. Kimura et al. (AIoTS 2022) investigated deep learning-based attacks on the small PRESENT-[4] block cipher with limited component changes, identifying characteristics specific to these attacks which remain unaffected by linear/differential cryptanalysis. Finding such characteristics is important because exploiting such characteristics can make the target cipher vulnerable to deep learning-based attacks. Thus, this paper extends a previous method to explore clues for designing symmetric-key cryptographic algorithms that are secure against deep learning-based attacks. We employ small PRESENT-[4] with two weak S-boxes, which are known to be weak against differential/linear attacks, to clarify the relationship between classical and deep learning-based attacks. As a result, we demonstrated the success probability of our deep learning-based whitebox analysis tends to be affected by the success probability of classical cryptanalysis methods. And we showed our whitebox analysis achieved the same attack capability as traditional methods even when the S-box of the target cipher was changed to a weak one.

Penetration testing is among the most efficient techniques to improve network system defense and search for potential weaknesses. Applying penetration testing with reinforcement learning can enhance automation and accuracy and reduce dependence on human labor. However, this approach still encounters obstacles in intricate network systems, such as large ones, where compromising is challenging. The lack of modeling derived from a specific common cybersecurity knowledge base also complicates effective applications in practice. Therefore, based on MITRE ATT&CK knowledge, we propose a multilayer action representation to improve the performance, accuracy, and applicability of penetration testing on complex networks. The multilayer action representation's goal is to embody actions in penetration testing as n-dimensional vectors while faithfully capturing their characteristics and relationships. Therefore, it directly improves the performance of reinforcement learning agents in large and complicated network scenarios. For faster training, we also use an epsilon-Wolpertinger architecture. We conducted experiments on four difficulty levels with three network configurations and 119 system scenarios and compared our approach with four different reinforcement learning techniques. Our approach not only represents and models actions with high accuracy but also improves the ability of reinforcement learning agents in a variety of difficult levels of network systems.

The increasing frequency and sophistication of cyberattacks makes it essential to keep up-to-date with threat information by using cyber threat intelligence (CTI). Structured CTI such as Structured Threat Information eXpression (STIX) is particularly useful because it can automate security operations such as updating FW/IDS rules and analyzing attack trends. However, as most CTIs are written in natural language, manual analysis with domain knowledge is required, which becomes quite time-consuming. In this work, we prose CyNER, a method for automatically structuring CTIs and converting them into STIX format. CyNER extracts named entities in the context of CTI and then extracts the relations between named entities and IOCs in order to convert them into STIX. In addition, by using key phrase extraction, CyNER can extract relations between IOCs that lack contextual information such as those listed at the bottom of a CTI, and named entities. We describe our design and implementation of CyNER and demonstrate that it can extract named entities with the F-measure of 0.80 and extract relations between named entities and IOCs with a maximum accuracy of 81.6%. Our analysis of structured CTI showed that CyNER can extract IOCs that are not included in existing reputation sites, and that it can automatically extract IOCs that have been exploited for a long time and across multiple attack groups. CyNER will therefore make CTI analysis more efficient.

There have been many vulnerabilities, and we need prompt countermeasures. One factor that makes more rapid measures necessary is Proof of Concept (PoC) codes. Although they are released to promote vulnerability countermeasures, attackers can also abuse them. In this paper, we analyze PoC codes that send HTTP requests, then generate IDS signatures. To analyze codes, there are two policies: dynamic analysis and static analysis. However, the former cannot cover the execution paths, and the latter cannot analyze dynamically determined values. In addition, symbolic execution compensates for their shortcomings, but its implementation cost is high. We propose a signature generation method for PoC codes that send HTTP requests based on an analysis combining dynamic and static analysis. We first statically explore execution paths of the code by searching for the conditional branch syntax using the abstract syntax tree. Then, we rewrite the branch conditions to enforce the specific execution path and generate a new code corresponding to each path. Finally, we execute each code, generate the attack requests dynamically, and extract signatures. The average detection rate for the requests was 86.9%. Moreover, we tested the signatures for 30 codes by actually executing them, and for nine codes, we detected the attack.

Fraudulent access by way of nInternet banking, credit cards and e-commerce are a serious problem. Fraudsters intend to steal credentials and log in to these websites in many ways such as phishing, malware infection, list based attack etc. There are products and services to prevent fraudulent access like fraud detection software and multi-factor authentication, however these have issues such as installation costs, detection accuracy and operation cost. Some security vendors provide client-side software to prevent fraud, but it is usually difficult for the companies to compel their end-users to install additional software because it may cause trouble and decrease usability. Regarding these issues we are researching an effective fraud detection method using server-side log information. In this paper, we show results from analyzing the attacker device attribute information and the environmental differences between genuine users and fraudsters based on the access log history from actual services and found that the attacker's environment changes year by year. We also discuss the effectiveness of the fraud detection methods described in previous research and effective detection methods utilizing real-world data.

Phishing via email has been spreading not only to individuals but also to companies, and various studies have been conducted on it. In addition, the use of chat has been increasing, and phishing via chat is assumed to be occurring, but the risk and susceptibility to phishing via chat have not been clarified. In this study, we conducted a questionnaire survey of 304 participants. In the survey, we divided the participants into two groups, designed similar phishing scenarios using email and chat, and conducted a role play susceptibility study. The results of the study indicated that the susceptibility of chat was as high as or higher than that of email, suggesting that phishing via chat is risky. On the basis of the results, we also summarized methods for reducing the risk of phishing in chat and recommendations for future research.

Modern Web services provide advanced features by utilizing hardware resources on the user's device. Web browsers implement a user consent-based permission model to protect user privacy. In this study, we developed P_ERMIUM, a web browser analysis framework that automatically analyzes the behavior of permission mechanisms implemented by various browsers. We systematically studied the behavior of permission mechanisms for 22 major browser implementations running on five different operating systems. We found fragmented implementations. Implementations between browsers running on different operating systems are not always identical. We determined that implementation inconsistencies could lead to privacy risks. We identified gaps between browser permission implementations and user perceptions from the user study corresponding to the analyses using P_ERMIUM. Based on the implementation inconsistencies, we developed two proof-of-concept attacks and evaluated their feasibility. The first attack uses permission information to secretly track the user. The second attack aims to create a situation in which the user cannot correctly determine the origin of the permission request and the user mistakenly grants permission. Finally, we clarify the technical issues that must be standardized in privacy mechanisms and provide recommendations to OS/browser vendors to mitigate the threats identified in this study.

Banks offering Online Banking services need to detect and prevent unauthorized electronic funds transfers to reduce financial crime risk. They monitor online banking transaction histories and use their own methods to detect and prevent unauthorized electronic fund transfers. However, unauthorized electronic fund transfers by criminals have not been eliminated. The average rate of false positives in the transaction monitoring systems installed in Japanese banks is up to 99%, indicating that the monitoring systems are not fully functional. Furthermore, the personnel responsible for fraud detection must manually check a large number of false positives, making it difficult for operators to be productive in their assigned tasks. Based on the above background, we develop a method to detect unauthorized electronic fund transfers and suspicious transactions with high accuracy using machine learning algorithms and evaluate its accuracy. Specifically, a supervised machine learning algorithm is applied to detect fraudulent transactions automatically. We evaluated the proposed method on a large set of online banking transaction data provided by a major Japanese bank for the period March 2019 to May 2020. We demonstrated that our approach could detect fraudulent activity with extremely high accuracy; FPR=0.000 and FNR=0.005 can be achieved for a security policy that minimizes false positives.

In this paper, we propose a new framework for security risk assessment. To conduct security analysis efficiently, it is necessary for developers to assess the security risks of machine learning based system (MLS) by themselves, but existing technologies cannot be used to such a purpose. Using the proposed framework, MLS developers can assess the security risks of MLSs by themselves. Our framework consists of two phases. In the preparation phase, a machine learning security expert extracts conditions of adversarial attacks for each adversarial attack method and makes an attack tree for each attack method using the extracted conditions. In addition, they prepare yes/no questions corresponding to extracted conditions. In the assessment phase, MLS developers just answer yes/no questions, and the assessment results are shown. We asked some developers to evaluate our proposal by implementing the proposed framework. As a result, they found some vulnerabilities in MLSs they chose to analyze. We received positive comments from them as results of the questionnaire.

Protecting a machine learning model and its inference inputs with secure computation is important for providing services with a valuable model. In this paper, we discuss how a model's parameter quantization works to protect the model and its inference inputs. To this end, we present an investigational protocol called MOTUS, based on ternary neural networks whose parameters are ternarized. Through extensive experiments with MOTUS, we found three key insights. First, ternary neural networks can avoid deterioration in accuracy due to secure computation with modulo operations. Second, the increment of model parameter candidates significantly improves accuracy more than an existing technique for accuracy improvement, i.e., batch normalization. Third, protecting both a model and inference inputs reduces inference throughput by four to seven times to provide the same level of accuracy compared with existing protocols protecting only inference inputs. We have released our source code via GitHub.

A README file plays an essential role as the face of a software project and the initial point of contact for developers in Open Source Software (OSS) projects. The code snippet ranks among the most important content in the README file for demonstrating the usage of software and APIs. While easy to comprehend, code snippets are preferred by clients in order to quickly understand software usage and features. However, proficient code snippets are sometimes found in README files. In this paper, we first investigate the prevalence of each competency level of Python code snippets in the README files. Then, we analyze the relationships between the usage of proficient code snippets and topics of libraries. From our empirical study on 1, 620 README files of PyPI libraries, we find that developers mainly present 92% of basic elements in code snippets. However, developers are likely to present proficient elements in code snippets from topics about Application Framework, Quality Assurance, and User Interface. We therefore (i) encourage developers to mainly present basic code snippets in their README files to attract more newcomers, and (ii) suggest that clients try to understand proficient code snippets if they are adopting libraries from previously mentioned topics.

Cross-industrial collaboration can yield disruptive innovations. However, there are fears regarding data sharing across different organizations. Even if data providers make contracts covering the data shared with their consumers, they will not be able to delete the shared data in accordance with the expiration dates and modify them after they were delivered to the consumers. Data consumers need to be extremely careful about management of shared data since huge penalties are imposed against violations of data protection laws. To allay these fears, we propose a system to handle external data management instead of data providers and consumers. In our system, containerized lifetime controllers delete expired shared data in accordance with contracts for shared data called life cycle policies. To allow only service programs stipulated in the policies to manipulate shared data, containerized volume controllers enforce the access control on the basis of the FUSE interceptions and the /proc file system. The proposed system is transparent to service programs because the containerized controllers run in execution environments that are separate from service programs. The proposed system can be applied to multiple container orchestration clusters in which a provider and consumer independently administer Kubernetes container orchestrators, as well as a single container orchestration cluster. We built a prototype system on Kubernetes container orchestrators presented by the Kubernetes community and public cloud service providers. Experimental results demonstrate that the proposed system achieves data sharing between a provider and consumer with moderate overheads for disk consumption of the containerized controllers, the extensions of the volume drivers, and execution time of the FUSE access control.

The RAM encryption encrypts the data on memory to prevent data leakage from an adversary to eavesdrop the memory space of the target program. The well-known implementation is Intel SGX, whose RAM encryption mechanism is definitely hardware dependent. In contrast, Watanabe et al. proposed a fully software-based RAM encryption scheme (SBRES). In this paper, we developed the tools for embedding the SBRES in C source codes for its practical application. We applied the tools to the source codes of some cryptographic implementations in Mbed TLS and confirmed that the tools successfully embedded the SBRES functionality in the cryptographic implementations.

A refinement type system can be used as a basis of a fully automated verification tool of higher-order functional programs. Most of existing refinement type systems are, however, designed for call-by-value programs, which are unsound for call-by-name programs. In this paper, we introduce a refinement type system for a call-by-name functional language. The most important difference between call-by-value and call-by-name for constructing a refinement type system is whether a variable can be treated as a value or not. The soundness of a typical refinement type systems depends on this fact. On the other hand, in a call-by-name program, since a variable is bound to a term, we cannot treat a variable as a value. To overcome this problem, we track when the term bound to a variable will be evaluated. If we found that a variable is already evaluated in some context, we can treat the variable as a value at the context. We also introduce a type inference algorithm and report on a prototype implementation and preliminary experiments.

We propose a fast algorithm called Z-TCA for triadic concept analysis (TCA). TCA is an extension of formal concept analysis (FCA), aiming at extracting ontologies by using mathematical order theories from a collection of ternary relations of three groups of variables: the object, attributes, and conditions. It finds various applications in fields like data mining and knowledge representation. However, the state-of-the-art TCA algorithms are suffering from the problem of low efficiency due to the complexity of the task. Attempts have been made to speed up the TCA process using a Binary Decision Diagram (BDD) or its improved version Zero-suppressed Decision Diagram (ZDD), while in this paper, we propose a new way to apply ZDD to TCA, named the Z-TCA algorithm. We conduct experiments on a real-world triadic context built from the IMDb database as well as some randomly-generated contexts and the results show that our Z-TCA algorithm can speed up the TCA process about 3 times compared to the baseline TRIAS algorithm. We also discover that when the density of the context exceeds 5%, our algorithm outperforms all other ZDD-based improved TCA algorithms and becomes the fastest choice for TCA.

We propose a content-based method for retrieving the 3D CAD assemblies which contain or are contained in the assembly given as a query from a database. The retrieval is based on ranking assemblies in the database according to their content rates which are different criteria from the similarities of shapes. The content rate is computed by comparing the projections of components constituting an assembly in the database with those of components constituting the query. We use the 3D Radon transform to obtain the projections. In existing methods for retrieving CAD models, projections of each model onto 2D planes, which are a set of two-dimensional data, are often used to compute the similarity between them. The proposed method simplifies the process of comparing the projections of components because the projections using the 3D Radon transform are a set of one-dimensional data. The method has the unique feature of identifying the component layouts, which reflect the technical know-how and information of the product designers. The components which have the same shapes but different properties such as material names are also distinguished based on their layouts without depending on the labels assigned to the properties. Other than our previous method, no existing methods possess such features. We show that the proposed method has better performance in the retrieval precision and processing time than the previous method.

NAND SSDs are widely used because of their improved performance and decreased cost. However, their performance behavior is complex. In this paper, we explore the relationship between the accumulation of operations after formatting and performance, and show mainly that writing to the unwritten area after formatting degrades read performance. As a further discussion, we also present the results of modeling read performance by approximating it as a quadratic function of two variables, and show that this modeling method can estimate performance.

Applying Differentially Private Stochastic Gradient Descent (DPSGD) to training modern, large-scale neural networks such as transformer-based models is a challenging task, as the magnitude of noise added to the gradients at each iteration scale with model dimension, hindering the learning capability significantly. We propose a unified framework, LSG, that fully exploits the low-rank and sparse structure of neural networks to reduce the dimension of gradient updates, and hence alleviate the negative impacts of DPSGD. The gradient updates are first approximated with a pair of low-rank matrices. Then, a novel strategy is utilized to sparsify the gradients, resulting in low-dimensional, less noisy updates that are yet capable of retaining the performance of neural networks. Empirical evaluation on natural language processing and computer vision tasks shows that our method outperforms other state-of-the-art baselines.

Recent technological advances in Virtual Reality (VR) and Augmented Reality (AR) enable users to experience a high-quality virtual world. The AR technology is attracting attention in various fields and is also used in the entertainment field such as museums. However, the existing AR technology generally requires specialized sensors such as Light Detection And Ranging (LiDAR) sensors and feature points, which require cost in terms of time and money. The authors have proposed a real-time background removal method and an AR system based on the estimated depth of the captured image to provide a virtual space experience using mobile devices such as smartphones. This paper describes an AR virtual space system that dynamically changes the replaced background based on motion information transmitted from the user's device.

With a range of learning platforms available on the Internet and the increasing use of video content in education, there have been many studies on using ontologies to support education and learning. However, in many cases, the connections among content systematized in this way are confined within a subject, and there is no learning across subjects or expansion of learners' interests beyond individual subjects. We aimed to connect all educational content in a way that goes beyond school subjects and curricula. In this study, we re-examined the learning sequence of subject matter terms over the course of the learning process, redefined it as a set of basic/advanced relationships among the terms, and proposed a method of estimating the appropriate ordering of content for learning by structuring and utilizing this terminological attribute. We also prototyped a system incorporating our proposed method to help users select educational content and tested it in a proof-of-concept experiment. By comparing the textbook-based presentation of related content with our proposed cross-subject presentation, we identified each method's characteristics and confirmed the proposed method's effectiveness at expanding learners' interests.

In this study, we presented MathTOUCH, a rich-text editor, to create mathematical documents. In mathematics e-learning with learning management systems, existing math input methods have heavy workloads; therefore, math notations are a long-standing challenge. Furthermore, when creating mathematical documents, such as writing reports and posting questions on the forum, users must enter both texts in a natural language input manner and mathematical expressions in a math input manner. Hence, users' workload increases. To address this issue, we developed the MathTOUCH editor to implement an intelligent math input interface that enables users to enter equations through predictive conversion from the colloquial-style text. A user study was conducted with 71 participants to evaluate the effectiveness of the proposed editor. The results indicate that users could enter equations with the MathTOUCH editor approximately 1.5 times faster than with a standard interface, and the participants also reported greater subjective satisfaction.

Vulnerable kernel code is a threat to an operating system kernel. An adversary's user process can forcefully invoke vulnerable kernel code to cause privilege escalation or denial of service (DoS). Although security engineers belong Computer Security Incident Response Team (CSIRT) and Product Security Incident Response Team (PSIRT) of service providers or security operators have to consider the effect of kernel vulnerability on their system environment when deciding whether or not to update a kernel, the list of vulnerable kernel code is not provided in the Common Vulnerabilities and Exposures (CVE) report In addition, it is difficult to identify the vulnerable kernel code from the exploitation results in the kernel which indicates the account or the kernel suspension information. To identify the details of kernel vulnerability, we propose a vulnerable kernel code tracer (vkTracer), which uses an alternative viewpoint using proof-of-concept (PoC) code to create a profile of kernel vulnerability. vkTracer traces the user process of the PoC code and the running kernel to hook the invocation of the vulnerable kernel code. Moreover, vkTracer extracts the whole kernel component's information using the running and static kernel image and debug section. This ensures that vkTracer identifies the virtual address range and function name of the invoked kernel code from the traced user process; a profile of kernel vulnerability is then created. The evaluation results indicate that vkTracer could trace PoC code execution (e.g., privilege escalation or DoS), identify vulnerable kernel code, and generate a kernel vulnerability profile. Additionally, vkTracer could trace the kernel code of system call invocation regarding CVE information. Furthermore, the implementation of vkTracer reveals that the identification overhead ranges from 5.2683 s to 5.2728 s on the PoC codes and the acceptable system call latency is 3.7197 µs. Moreover, vkTracer represents 0.37% and 0.56% of the dynamic kernel tracing overhead for the web client program access overhead of 100,000 Hypertext Transfer Protocol sessions.

Cyber attacks are causing tremendous damage around the world. To protect against attacks, many organizations have established or outsourced Security Operation Centers (SOCs) to check a large number of logs daily. Since there is no perfect countermeasure against cyber attacks, it is necessary to detect signs of intrusion quickly to mitigate damage caused by them. However, it is challenging to analyze a lot of logs obtained from PCs and servers inside an organization. Therefore, there is a need for a method of efficiently analyzing logs. In this paper, we propose a recommendation system using the ATT&CK technique, which predicts and visualizes attackers' behaviors using collaborative filtering so that security analysts can analyze logs efficiently. We evaluated the proposed method using real-world cyber-attack cases and found that it is able to make predictions with higher recall than our previously proposed method.

Privacy protection with synthetic data generation often uses differentially private statistics and model parameters to quantitatively express theoretical security. However, these methods do not take into account privacy protection due to the randomness of data generation. In this paper, we theoretically evaluate Rényi differential privacy of the randomness in data generation of a synthetic data generation method that uses the mean vector and the covariance matrix of an original dataset. Specifically, for a fixed α > 1, we show the condition of ϵ such that the synthetic data generation satisfies (α, ϵ)-Rényi differential privacy under a bounded neighboring condition and an unbounded neighboring condition, respectively. In particular, under the unbounded condition, when the size of the original dataset and synthetic dataset is 10 million, the mechanism satisfies (4, 0.576)-Rényi differential privacy. We also show that when we translate it into the traditional (ϵ, δ)-differential privacy, the mechanism satisfies (4.46, 10^-14)-differential privacy.

The local differential privacy metric has attracted attention due to its quantitative nature, and many mechanisms have been studied for satisfying local differential privacy based on data formats and use cases. Local differential privacy mechanisms generally target a certain data space and perturb it sufficiently to provide indistinguishability of the data on that space. Therefore, individual data tends to be greatly disturbed so that even relatively simple tasks require a large amount of data to equalize the noise caused by the mechanism. In this paper, we define hierarchical local differential privacy, which is an extension of local differential privacy, and propose a mechanism to satisfy both local differential privacy and hierarchical local differential privacy. Hierarchical local differential privacy views a data space hierarchically as a set of smaller spaces, and instead of abandoning the privacy of data contained in different spaces, the amount of noise can be reduced. In this paper, we further design a hierarchical local differential privacy framework and achieve a privacy guarantee based on local differential privacy for all the data in the framework. Finally, we experimentally evaluate the proposed framework using image data. The framework allows control over the amount of information that can be disclosed, and furthermore, maintains a higher degree of utility than applying a simple local differential privacy mechanism.

Companies and organizations inform users of how they handle personal data through privacy policies on their websites. Particular information, such as the purposes of collecting personal data and what data are provided to third parties is required to be disclosed by laws and regulations. An example of such a law is the Act on the Protection of Personal Information in Japan. In addition to privacy policies, an increasing number of companies are publishing security policies to express compliance and transparency of corporate behavior. However, it is challenging to update these policies against legal requirements due to the periodic law revisions and rapid business changes. In this study, we developed a method for analyzing privacy policies to check whether companies comply with legal requirements. In particular, the proposed method classifies policy contents using bidirectional encoder representations from transformers and evaluates privacy compliance by comparing the classification results with legal requirements. In addition, we analyzed security policies using the proposed method, to confirm whether the combination of privacy and security policies contributes to privacy compliance. In this study, we collected and evaluated 1,298 privacy policies and 139 security policies for Japanese companies. The results revealed that over 90% privacy policies adequately describe the handling of personal information by first parties, user rights, and security measures, and over 90% insufficiently describe the data retention and specific audience. These differences in the number of descriptions depend on industry guidelines and business characteristics. Additionally, security policies were found to improve the compliance rates of 48 out of 139 companies by describing security practices not included in privacy policies. Finally, we conducted a comparative analysis of policies before and after the Japanese law revision. We identified that although companies updated their policies, these updates were insufficient to meet the new law requirements.

Software supply chain attacks have become more prevalent due to attacker's higher skills and sophisticated attack methods. These attacks can go unnoticed for a long time and cause widespread damage. Therefore, monitoring all software products in the supply chain to detect these attacks is crucial. However, the software supply chain is complex, and tracking suspicious changes to third-party products, services, and libraries used to build software products can be challenging. This makes it difficult for developers to identify and confirm vulnerabilities in the software. Even one vulnerable file in the supply chain can have severe consequences if exploited, so regular vulnerability checks are necessary to keep software products safe. To address this issue, we propose a Vulnerability Assessment Tool that identifies newly found vulnerabilities from public databases and efficiently checks them in existing internal products. Our method uses a two-step process. The first step is estimating related products and directly affected products for a given CVE ID. The second step is a detailed check of the vulnerabilities in the system based on the results estimated in Step 1. This assessment method reduces the time the PSIRT Analyst requires to assess the vulnerabilities.

To handle high-resolution images on finite computational resources, many researches have been conducted on hierarchical networks to load features in only the most meaningful local regions. However, it is difficult to determine the correct number and location of patch regions at the appropriate scale in these methods. Then, incorrectly selected regions at different scales interfere with feature extraction and information integration. To solve this issue, we propose a hierarchical attention weighted network (HAWK-Net), which consists of a backbone network with differentiable Top-K selection and spatially gated blocks. The Top-K selected patches are identified from multiple image scaled features and extracted from an original high-resolution image. Then, patch features are aggregated via a novel gate mechanism under the uncertainty of the predicted information. Not only can multi-scale information uncertainty be modeled, but it also controls the gradient to the feature network coming from patch images with low confidence in the region proposal network in feedback during training. Our model is a simple yet efficient network structure that can learn from multiple scales and patches and is capable of end-to-end training. Based on benchmarks of multiple high-resolution images, our model achieves even higher performance with lower memory usage and reduced computation time.

Mixed precision numerical methods using low precision computing have attracted much attention under recent computational hardware trends. In this research, we focus on solving large, sparse, and non-symmetric linear systems, and consider developing a numerical method based on a mixed precision variant of the iterative refinement scheme (MP-IR), in which we can exploit low precision computing and provide a computed solution with the same accuracy as that obtained by conventional methods without low precision computing. We employ the BiCGSTAB solver with FP32 as an inner solver of MP-IR and investigate its numerical behavior through numerical experiments. From the analyses on the obtained results including a comparison with MP-IR using GMRES(m), which is also known as MP-GMRES(m) and has been widely studied, the potential of MP-IR using BiCGSTAB has been confirmed. Together with other obtained results, this paper provides insights that are helpful in developing an efficient mixed precision linear solver for practical applications.

This paper focuses on a solution process of a sequence of linear systems having an asymmetric coefficient matrix. We propose a new subspace correction (SC) preconditioning method based on error vector sampling for asymmetric coefficient matrices. In the method, we identify approximate left and right singular vectors associated with small singular values using the sampled error vectors calculated after the first solution step. Using these singular vectors, we construct auxiliary matrices to specify the subspace used in the SC preconditioning. We applied the proposed preconditioning method to an ILU(0)-Bi-CGSTAB solver in additive Schwarz manner. To examine the proposed preconditioning method, we conducted numerical tests using matrices downloaded from the SuiteSparse Matrix Collection database. Numerical tests showed that the proposed method could reduce the number of iterations to convergence and the solution time in many test cases, which confirms the effectiveness of the method. Moreover, the proposed method succeeded to get convergence for some test problems that the ILU(0) preconditioned solver could not solve.