A5070030398- Anomaly Detection Techniques and Applications
- Adversarial Robustness in Machine Learning
- Network Security and Intrusion Detection
- Time Series Analysis and Forecasting
- Advanced Neural Network Applications
- Machine Learning and Data Classification
- Topic Modeling
- Advanced Malware Detection Techniques
- Generative Adversarial Networks and Image Synthesis
- Advanced Graph Neural Networks
- Recommender Systems and Techniques
- Natural Language Processing Techniques
- Domain Adaptation and Few-Shot Learning
- Digital Media Forensic Detection
- Privacy-Preserving Technologies in Data
- Data Stream Mining Techniques
- Music and Audio Processing
- Speech and dialogue systems
- Data Management and Algorithms
- Advanced Clustering Algorithms Research
- Complex Systems and Time Series Analysis
- Context-Aware Activity Recognition Systems
- Neural Networks and Applications
- Human Mobility and Location-Based Analysis
- Physical Unclonable Functions (PUFs) and Hardware Security
The University of Melbourne
2016-2025
Fukushima Prefectural Culture Center
2024
Information Technology University
2023
Deakin University
2021
Lampung University
2021
John Wiley & Sons (United States)
2019
Intelligent Systems Research (United States)
2019
Data61
2014-2016
Bounding box (bbox) regression is a fundamental task in computer vision. So far, the most commonly used loss functions for bbox are Intersection over Union (IoU) and its variants. In this paper, we generalize existing IoU-based losses to new family of power IoU that have term an additional regularization with single parameter $\alpha$. We call $\alpha$-IoU analyze properties such as order preservingness loss/gradient reweighting. Experiments on multiple object detection benchmarks models...
Datasets with significant proportions of noisy (incorrect) class labels present challenges for training accurate Deep Neural Networks (DNNs). We propose a new perspective understanding DNN generalization such datasets, by investigating the dimensionality deep representation subspace samples. show that from perspective, DNNs exhibit quite distinctive learning styles when trained clean versus proportion labels. Based on this finding, we develop dimensionality-driven strategy, which monitors...
Robust loss functions are essential for training accurate deep neural networks (DNNs) in the presence of noisy (incorrect) labels. It has been shown that commonly used Cross Entropy (CE) is not robust to Whilst new have designed, they only partially robust. In this paper, we theoretically show by applying a simple normalization that: any can be made However, practice, simply being sufficient function train DNNs. By investigating several functions, find suffer from problem underfitting. To...
Feature interactions are essential for achieving high accuracy in recommender systems. Many studies take into account the interaction between every pair of features. However, this is suboptimal because some feature may not be that relevant to recommendation result and taking them introduce noise decrease accuracy. To make best out interactions, we propose a graph neural network approach effectively model them, together with novel technique automatically detect those beneficial terms The...
Quantum machine learning (QML) has received increasing attention due to its potential outperform classical methods in problems such as classification and identification tasks. A subclass of QML is quantum generative adversarial networks (QGANs) which have been studied a counterpart GANs widely used image manipulation generation The existing work on QGANs still limited small-scale proof-of-concept examples based images with significant down-scaling. Here we integrate techniques propose new...
Machine learning (ML) methods such as artificial neural networks are rapidly becoming ubiquitous in modern science, technology and industry. Despite their accuracy sophistication, can be easily fooled by carefully designed malicious inputs known adversarial attacks. While vulnerabilities remain a serious challenge for classical networks, the extent of existence is not fully understood quantum ML setting. In this work, we benchmark robustness variational classifiers (QVC), at scale performing...
One of the most crucial tasks for utility companies is load forecasting in order to plan future demand generation capacity and infrastructure. Improving accuracy over a short period challenging open problem due variety factors that influence load, volume data needs be considered. This paper proposes new approach term using an effective combination clustering deep learning methods, along with weighted aggregation mechanism. Our evaluation smart meter from publicly available real-life dataset...
Deep Neural Networks (DNNs) have recently been shown to be vulnerable against adversarial examples, which are carefully crafted instances that can mislead DNNs make errors during prediction. To better understand such attacks, a characterization is needed of the properties regions (the so-called 'adversarial subspaces') in examples lie. We tackle this challenge by characterizing dimensional regions, via use Local Intrinsic Dimensionality (LID). LID assesses space-filling capability region...
Random projection is a popular method for dimensionality reduction due to its simplicity and efficiency. In the past few years, random fuzzy c-means based cluster ensemble approaches have been developed high-dimensional data clustering. However, they require large amounts of space storing big affinity matrix, incur computation time while clustering in this matrix. paper, we propose new projection, framework data. Our uses cumulative agreement aggregate partitions. Fuzzy partitions...
One-class support vector machines (OCSVMs) are very effective for semisupervised anomaly detection. However, their performance strongly depends on the settings of hyperparameters, which has not been well studied. Moreover, unavailability a clean training set that only comprises normal data in many real-life problems given rise to application OCSVMs an unsupervised manner. it shown if includes anomalies, boundary created by is prone skew toward anomalies. This problem decreases detection rate...
Protecting personal data against exploitation of machine learning models is crucial. Recently, availability attacks have shown great promise to provide an extra layer protection the unauthorized use train neural networks. These methods aim add imperceptible noise clean so that networks cannot extract meaningful patterns from protected data, claiming they can make "unexploitable." This paper provides a strong countermeasure such approaches, showing unexploitable might only be illusion. In...
High-fidelity digital human representations are increasingly in demand the world, particularly for interactive telepresence, AR/VR, 3D graphics, and rapidly evolving metaverse. Even though they work well small spaces, conventional methods reconstructing motion frequently require use of expensive hardware have high processing costs. This study presents HumanAvatar, an innovative approach that efficiently reconstructs precise avatars from monocular video sources. At core our methodology, we...
End-to-end task-oriented dialogue systems aim to generate system responses directly from plain text inputs. There are two challenges for such systems: one is how effectively incorporate external knowledge bases (KBs) into the learning framework; other accurately capture semantics of history. In this paper, we address these by exploiting graph structural information in base and dependency parsing tree dialogue. To leverage history, propose a new recurrent cell architecture which allows...
Detecting beneficial feature interactions is essential in recommender systems, and existing approaches achieve this by examining all the possible interactions. However, cost of higher-order prohibitive (exponentially growing with order increasing). Hence only detect limited (e.g., combinations up to four features) interactions, which may miss orders higher than limitation. In paper, we propose a hypergraph neural network based model named HIRS. HIRS first work that directly generates...
Contrastive language-image pretraining (CLIP) has been found to be vulnerable poisoning backdoor attacks where the adversary can achieve an almost perfect attack success rate on CLIP models by only 0.01\% of training dataset. This raises security concerns current practice large-scale unscrutinized web data using CLIP. In this work, we analyze representations backdoor-poisoned samples learned and find that they exhibit unique characteristics in their local subspace, i.e., neighborhoods are...
The rapid advancement of large models, driven by their exceptional abilities in learning and generalization through large-scale pre-training, has reshaped the landscape Artificial Intelligence (AI). These models are now foundational to a wide range applications, including conversational AI, recommendation systems, autonomous driving, content generation, medical diagnostics, scientific discovery. However, widespread deployment also exposes them significant safety risks, raising concerns about...