Institute for Information Industry

As network communication technologies rapidly advance, ransomware has emerged as a significant cybersecurity threat that organizations cannot ignore. Static analysis enables rapid identification of ransomware by examining file structure and code characteristics before execution. However, existing classifiers are predominantly designed under the closed-set assumption, causing them to misclassify novel variants into known families. Furthermore, ransomware datasets typically exhibit long-tailed distributions with emerging families having very few available samples, making it difficult for models to learn discriminative features. To address these challenges, we propose Few-Shot Open-Set Ransomware Detection through Meta-learning and Energy-based Modeling (MEM), a unified open-set recognition framework based on static analysis of Portable Executable features. By integrating Model-agnostic Meta-learning (MAML), the model rapidly adapts to new families with limited samples. The Energy Function quantifies the confidence of predictions in distinguishing between known samples and unknown ones, while Focal Loss dynamically adjusts sample weights to reduce bias introduced by imbalanced distributions. The experimental results demonstrate that MEM achieves higher classification accuracy and better rejection performance of unknown samples than existing open-set recognition methods.

There are apps for everything, from online banking, social software to shopping. They have become one of the most important tools in our daily lives. This even implies that a mobile device has stored most of personal information, including photos, credit cards, and communications. If an intruder succeeds in hacking into the mobile device, all private properties must suffer from the leakage threats. Undoubtedly, a malware is the commonest tool used by an attacker to compromise a mobile phone. In particular, it is often disguised as a popular application through an obfuscated or packed form. That is the main reason why it is difficult to distinguish a malware from the legal ones. In this article, we have adopted machine learning technique to develop a static analysis mechanism for Android malware detection based on shallow feature and permission correlation (AMD). AMD first analyses the Application Programming Interfaces of the target to detect all possible and hidden privilege threats. It then filters this obfuscation information using permission correlation to eliminate noise and identify meaningful malicious indicators. The proposed approach leverages the correlation patterns between permissions and API calls to distinguish suspicious behaviours from legitimate ones. Thus, AMD can extract all the representative shallow features to achieve the high detection rate. Simulation results have shown that AMD can outperform related works under the datasets of CICAndMal2017 and CICMalDroid2020, which confirms the effectiveness of shallow features and permission correlation.

International Telecommunication Union clearly defines IoT (Internet of Things) as “a phenomenon where every standalone object is connected”, this shows that as connectivity and communications technology gradually mature, various linking devices and equipment will be able to connect and create a more rich and diverse service scenario. However, as this trend is being established, these connected devices will produce large amount of data, which in turn will cause two problems that need to be solved: first, the increasing number of connected devices will lead to data congestion. The current solution is to process data in batches to divert data flow, but this will still result in latency and inability to respond to the issue immediately; second, with the advent of more wireless and cable-connected devices, service operators will have to invest more into the installation of cloud infrastructure. Aside from the initial installation, follow-up maintenance will also require huge costs, bumping up service prices. In response to the problems with the IoT, the structure of edge computing called “fog computing” was created. It can be defined as “an emerging method of computing with the main goal of shifting computing and connecting units from the cloud to the edge or nearedge position to help devices preemptively and autonomously process information and data created by connected devices, creating a computing structure that can operate without relying on the center.

The chapter “AI Safety and Security” presents a comprehensive and multi-dimensional exploration, addressing the critical aspects of safety and security in the context of large language models. The chapter begins by identifying the risks and threats posed by LLMs, delving into vulnerabilities such as bias, misinformation, and unintended AI interactions, impacts like privacy concerns. Building on these identified risks, it then explores the strategies and methodologies for ensuring AI safety, focusing on principles like robustness, transparency, and accountability and discussing the challenges of implementing these safety measures. It concludes with an insight into long-term AI safety research, highlighting ongoing efforts and future directions to sustain AI system safety amidst rapid technological advancements and encouraging a collaborative approach among various stakeholders. By integrating perspectives from computer science, ethics, law, and social sciences, the chapter provides an insightful and comprehensive analysis of current and future challenges in AI safety and security.

Existing research has either adapted the Probably Approximately Correct (PAC) Bayesian framework for federated learning (FL) or used information-theoretic PAC-Bayesian bounds while introducing their theorems, but few consider the non-IID challenges in FL. Our work presents the first non-vacuous federated PAC-Bayesian bound tailored for non-IID local data. This bound assumes unique prior knowledge for each client and variable aggregation weights. We also introduce an objective function and an innovative Gibbs-based algorithm for the optimization of the derived bound. The results are validated on real-world datasets.

In the post-pandemic era, AI development has entered a new phase, driven by the emergence of generative AI and the increasing prevalence of information and communication technologies represented by 5G. In the dawn of the AI age, significant shifts are underway, heralding a new era of transformation. This article explores the impact of AI advancement on Taiwan's legal framework during this era. It delves into both historical developments and prospective transformations and begins with a retrospective analysis of Taiwan's AI-related legislation before the post-pandemic era. Subsequently, it focuses on a critical aspect of Taiwan's AI legal framework during this era, including two pivotal documents in Taiwan: the "AI Basic Act" and the " Reference Guidelines for the Use of Generative AI in the Executive Yuan and its Subordinate Agencies and Affiliated Institutions”. Both have recently garnered widespread attention, marking a crucial step towards navigating the evolving landscape of artificial intelligence. Finally, this article offers insights into the prospective evolution of Taiwan's legal landscape pertaining to AI. The formulation of the "AI Basic Act" is pivotal in setting fundamental norms for AI development in Taiwan. This draft is significant with its 24 articles aimed at fostering a trustworthy environment and promoting diverse AI utilization. It ensures that Taiwan's high-tech advancements align with global standards. On the other side, the “Reference Guidelines for the Use of Generative AI” are designed to regulate the administrative agencies in response to the widespread application of generative AI. Furthermore, since 2023, the Personal Data Protection Act and Criminal Code, among others, have all been amended to include provisions related to AI technology, reflecting the apparent impact of AI on the legal system. Given the novelty of AI in society, prioritizing the creation of a basic act is crucial. Such legislation defines AI in Taiwan, establishes ethical principles, outline development directions, and provides fundamental norms. It will serve as a cornerstone for shaping and revising other policies and laws to navigate the evolving landscape of AI.

In the 5G massive machine type communication (mMTC) scenario, user equipment with poor signal quality requires numerous repetitions to compensate for the additional signal attenuation. However, an excessive number of repetitions consumes additional wireless resources, decreasing the transmission rate, and increasing the energy consumption. An insufficient number of repetitions prevents the successful deciphering of the data by the receivers, leading to a high bit error rate. The present study developed adaptive repetition approaches with the k-nearest neighbor (KNN) and support vector machine (SVM) to substantially increase network transmission efficacy for the enhanced machine type communication (eMTC) system in the 5G mMTC scenario. The simulation results showed that the proposed repetition with the learning approach effectively improved the probability of successful transmission, the resource utilization, the average number of repetitions, and the average energy consumption. It is therefore more suitable for the eMTC system in the mMTC scenario than the common lookup table.

The Inframammary fold is the lower breast boundary connected to the underlying chest wall. The presence of an inframammary fold is one of the positioning criteria for mammography quality evaluation. As the inframammary fold appeared on a MedioLateral Oblique- view mammogram means that posterior breast tissue is well-presented. This paper introduces an inframammary fold classification method based on a convolutional neural network (CNN) for mammography quality evaluation. The classifier utilizes three CNN models separately, including, AlexNet, GoogLeNet, and Vgg-16. In addition, we implement the two-category and the three-category classifications. Our method extracts a class activation map and then analyzes it to interpret what features the classifier observes. The experiment results demonstrate that the proposed method outperforms the previous work. The CAM depicts why the classifier efficiently determines the presence of an inframammary fold.

Smart agriculture is introduced with the latest technologies to improve the overall agricultural production capacity and efficiency. This paper proposes a quantum-inspired decision making scheme for crop cultivation in smart agriculture. The goal of the scheme is to increase the farmers’ welfare by planting different types of crop on adjacent farmlands to achieve optimal balance of crop demand and production. The proposed scheme is modeled as a graph coloring problem, which is computationally hard to solve for large and complex graphs in general. By taking advantage of quantum-inspired metaheuristic approach, the graph coloring problem is formulated as Quadratic Unconstrained Binary Optimization (QUBO) model and solved by a quantum annealer. The experimental realization of the proposed method is investigated on D-Wave quantum annealing devices and its performance is compared with those of the brute-force and greedy algorithms. The experimental results show that the proposed decision making scheme achieves a higher success probability than the greedy algorithm and a much lower run time than the brute-force approach.

For machine learning in online social networks or mobile environments, data privacy-preserving is a very important issue. In the past, artificial intelligence and machine learning algorithms that operating on a stand-alone machine must analyze each data in order to build an accurate model when training a model. However, this method is insufficient in terms of privacy protection, because it must collect data from different sources; on the other hand, if these algorithms analyze only a part of the data, the accuracy of the models constructed by artificial intelligence and machine learning algorithms will be very low. To solve the above problems, this study starts from the privacy protection aspect of industrial data management, combined with the concept of federated learning, and hopes to improve the accuracy of algorithm modules while ensuring data privacy. Therefore, in this paper, we analyze the impact of federated learning under different protocol mechanisms on the accuracy of algorithm modules, and based on this, we can apply this module to integrate blockchain smart contract technology in the future.