A5101547684- Adversarial Robustness in Machine Learning
- Anomaly Detection Techniques and Applications
- Advanced Neural Network Applications
- Second Language Learning and Teaching
- Impact of Technology on Adolescents
- Traffic Prediction and Management Techniques
- Explainable Artificial Intelligence (XAI)
- Transportation Planning and Optimization
- Knowledge Management and Sharing
- Digital Marketing and Social Media
- Foreign Language Teaching Methods
- Translation Studies and Practices
- Physical Unclonable Functions (PUFs) and Hardware Security
- Domain Adaptation and Few-Shot Learning
- Social Media and Politics
- Organizational Learning and Leadership
- Language, Communication, and Linguistic Studies
- Advanced Optimization Algorithms Research
- Advanced Malware Detection Techniques
- Gender Studies in Language
- Evaluation Methods in Various Fields
- Advancements in Semiconductor Devices and Circuit Design
- EFL/ESL Teaching and Learning
- Social Media in Health Education
- COVID-19 diagnosis using AI
Hong Kong University of Science and Technology
2018-2025
University of Hong Kong
2018-2025
Education University of Hong Kong
2023-2024
Tencent (China)
2023
University of Science and Technology of China
2014
Tsinghua University
2013
We present a new method for black-box adversarial attack. Unlike previous methods that combined transfer-based and scored-based by using the gradient or initialization of surrogate white-box model, this tries to learn low-dimensional embedding pretrained then performs efficient search within space attack an unknown target network. The produces perturbations with high level semantic patterns are easily transferable. show approach can greatly improve query efficiency across different network...
Zeroth-order optimization algorithms recently emerge as a popular research theme in and machine learning, playing important roles many deep-learning related tasks such black-box adversarial attack, deep reinforcement well hyper-parameter tuning. Mainstream zeroth-order algorithms, however, concentrate on exploiting zeroth-order-estimated first-order gradient information of the objective landscape. In this paper, we propose novel meta-algorithm called Hessian-Aware Zeroth-Order (ZOHA)...
While adversarial training and its variants have shown to be the most effective algorithms defend against attacks, their extremely slow process makes it hard scale large datasets like ImageNet. The key idea of recent works accelerate is substitute multi-step attacks (e.g., PGD) with single-step FGSM). However, these methods suffer from catastrophic overfitting, where accuracy PGD attack suddenly drops nearly 0% during training, network totally loses robustness. In this work, we study...
Zeroth-order optimization is an important research topic in machine learning. In recent years, it has become a key tool black-box adversarial attack to neural network based image classifiers. However, existing zeroth-order algorithms rarely extract second-order information of the model function. this paper, we utilize objective function and propose novel \textit{Hessian-aware algorithm} called \texttt{ZO-HessAware}. Our theoretical result shows that \texttt{ZO-HessAware} improved convergence...
More and more college students are using microblogs, with some excessive users demonstrating addiction-like symptoms. However, there is currently no published scale available for use in assessing of these a significant impediment to advancing this area research. We collected data from 3,047 China developed Microblog Excessive Use Scale (MEUS) Chinese students, comparing it criteria used Internet addiction. Our diagnostic featured three factors, two which–"withdrawal health problem" "time...
We present a new method for score-based adversarial attack, where the attacker queries loss-oracle of target model. Our employs parameterized search space with structure that captures relationship gradient loss function. show searching over structured can be approximated by time-varying contextual bandits problem, takes feature associated arm to make modifications input, and receives an immediate reward as reduction The problem then solved Bayesian optimization procedure, which take...
Various social and cultural factors have influenced language as a communicative carrier of human society. As form communication between diverse languages cultures, translation is process transforming symbols transplantation abutment. To facilitate exchanges China other countries while serving cross-cultural communication, it great significance for translators to deeply comprehend the discrepancies artistic images source target language, choose appropriate strategies, correctly deal with...
With the development of globalization and surging demand for foreign language talents, it is particularly significant to foster intercultural communicative awareness enhance competence. Undoubtedly, English reading serves as a crucial part learning, improving ability also one important ways cultivate students' cross-cultural However, both linguistic non-linguistic differences between Chinese create barriers reading. Therefore, this paper probes into training strategies college from...
Adversarial training is a popular method to robustify models against adversarial attacks. However, it exhibits much more severe overfitting than on clean inputs. In this work, we investigate phenomenon from the perspective of instances, i.e., input-target pairs. Based quantitative metric measuring instances' difficulty, analyze model's behavior instances different difficulty levels. This lets us show that decay in generalization performance result attempt fit hard instances. We theoretically...
A period of construction developed rapidly has already came up in urban public transport system, and which rail transit is the backbone. Urban consists light with large capacity subway. The trend transportation's development division labor cooperation all modes transportation. Then a model traffic integration could be formed. joins between various transportation are important links. Joins coordinators system other means include two points, that links paths transfers stations. This paper...
While adversarial training and its variants have shown to be the most effective algorithms defend against attacks, their extremely slow process makes it hard scale large datasets like ImageNet. The key idea of recent works accelerate is substitute multi-step attacks (e.g., PGD) with single-step FGSM). However, these methods suffer from catastrophic overfitting, where accuracy PGD attack suddenly drops nearly 0% during training, destroying robustness networks. In this work, we study...