A5056896393- Adversarial Robustness in Machine Learning
- Topic Modeling
- Advanced Malware Detection Techniques
- Anomaly Detection Techniques and Applications
- Reinforcement Learning in Robotics
- Explainable Artificial Intelligence (XAI)
- Landslides and related hazards
- Natural Language Processing Techniques
- Multimedia Communication and Technology
- Scheduling and Optimization Algorithms
- Peer-to-Peer Network Technologies
- Advanced machining processes and optimization
- Complex Network Analysis Techniques
- Online Learning and Analytics
- Robotics and Sensor-Based Localization
- Computational Geometry and Mesh Generation
- Educational Technology and Assessment
- Video Analysis and Summarization
- Image Processing Techniques and Applications
- Advanced Surface Polishing Techniques
- Neural Networks and Applications
- Tree Root and Stability Studies
- Mobile and Web Applications
- Machine Learning in Materials Science
- Advanced Graph Neural Networks
Sichuan University of Arts and Science
2024
AviChina Industry & Technology (China)
2023
Planetary Science Institute
2023
University of California, Los Angeles
2023
National University of Defense Technology
2019-2022
Huawei Technologies (Sweden)
2021
Wyoming Department of Education
2001
Abstract Landslides are notoriously difficult to predict because numerous spatially and temporally varying factors contribute slope stability. Artificial neural networks (ANN) have been shown improve prediction accuracy but largely uninterpretable. Here we introduce an additive ANN optimization framework assess landslide susceptibility, as well dataset division outcome interpretation techniques. We refer our approach, which features full interpretability, high accuracy, generalizability low...
The utilisation of foundation models as smartphone assistants, termed app agents, is a critical research challenge. These agents aim to execute human instructions on smartphones by interpreting textual and performing actions via the device's interface. While promising, current approaches face significant limitations. Methods that use large proprietary models, such GPT-4o, are computationally expensive, while those smaller fine-tuned often lack adaptability out-of-distribution tasks. In this...
We present a framework for intuitive robot programming by non-experts, leveraging natural language prompts and contextual information from the Robot Operating System (ROS). Our system integrates large models (LLMs), enabling non-experts to articulate task requirements through chat interface. Key features of include: integration ROS with an AI agent connected plethora open-source commercial LLMs, automatic extraction behavior LLM output execution actions/services, support three modes...
Deep neural networks (DNNs) have achieved remarkable success in various tasks such as image classification, speech recognition, and natural language processing. However, DNNs proven to be vulnerable attacks from adversarial examples. These samples are generated by adding some imperceptible disturbances, which used mislead the output decision of deep learning model bring significant security risks system. previous research mainly focused on computer vision, thus neglecting issues processing...
It remains important to make abnormity detection from largescale behavioral data of Internet. Existing related approaches mostly failed employ high-dimensional characteristics Internet data, which limits the effect. To deal with this issue, we introduce graph convolution network (GCN) generate fine-grained feature representation towards data. And a deep GCN-based model for is proposed in paper. Firstly, GCN used extract global co-occurrence information behavior Then, embedding applied...
On-device control agents, especially on mobile devices, are responsible for operating devices to fulfill users' requests, enabling seamless and intuitive interactions. Integrating Multimodal Large Language Models (MLLMs) into these agents enhances their ability understand execute complex commands, thereby improving user experience. However, fine-tuning MLLMs on-device presents significant challenges due limited data availability inefficient online training processes. This paper introduces...
Smartphone agents are increasingly important for helping users control devices efficiently, with (Multimodal) Large Language Model (MLLM)-based approaches emerging as key contenders. Fairly comparing these is essential but challenging, requiring a varied task scope, the integration of different implementations, and generalisable evaluation pipeline to assess their strengths weaknesses. In this paper, we present SPA-Bench, comprehensive SmartPhone Agent Benchmark designed evaluate...
This paper introduces a novel mobile phone control architecture, termed ``app agents", for efficient interactions and controls across various Android apps. The proposed Lightweight Multi-modal App Control (LiMAC) takes as input textual goal sequence of past observations, such screenshots corresponding UI trees, to generate precise actions. To address the computational constraints inherent smartphones, within LiMAC, we introduce small Action Transformer (AcT) integrated with fine-tuned...
Adversarial examples reveal the fragility of deep learning models. Recent studies have shown that models are also vulnerable to universal adversarial perturbations. When input-agnostic sequence words concatenated any input instance in data set, it fools model produce a specific prediction [9] and [10]. Despite being highly successful, they often need obtain gradient information target model. However, under more realistic black box conditions, we can only manipulate output model, which brings...
Deep learning models are vulnerable to backdoor attacks. The success rate of textual attacks based on data poisoning in existing research is as high 100%. In order enhance the natural language processing model’s defense against attacks, we propose a method via poisoned sample recognition. Our consists two parts: first step add controlled noise layer after model embedding layer, and train preliminary with incomplete or no embedding, which reduces effectiveness samples. Then, use initially...
A key method for creating Artificial Intelligence (AI) agents is Reinforcement Learning (RL). However, constructing a standalone RL policy that maps perception to action directly encounters severe problems, chief among them being its lack of generality across multiple tasks and the need large amount training data. The leading cause it cannot effectively integrate prior information into perception-action cycle when devising policy. Large language models (LLMs) emerged as fundamental way...
In order to achieve the justification of whether a frame structure detection exists or not under condition that protocol is unknown, an algorithm for in physical layer based on image proposed. Firstly, sequence arranged into analysis matrix and converted image. Then, according characteristics frames, it deduced corresponding images have texture features, is, black white blocks appear alternately. Finally, support vector machine classifier used detect structure. Simulation results show...
Relation extraction is a fundamental task in natural language processing, aiming at extracting relational triples from plain text. However, there are fewer instances the manually constructed dataset to meet learning needs of relation models. Distant supervision approach has attracted interest numerous researchers due its ability construct large datasets low cost. Nevertheless, certain problems with distant overly strong assumptions. In this paper, we introduce three main supervised...
Despite deep neural networks (DNNs) having achieved impressive performance in various domains, it has been revealed that DNNs are vulnerable the face of adversarial examples, which maliciously crafted by adding human-imperceptible perturbations to an original sample cause wrong output DNNs. Encouraged numerous researches on examples for computer vision, there growing interest designing attacks Natural Language Processing (NLP) tasks. However, attacking NLP is challenging because text...
Recent studies have shown that natural language processing (NLP) models are vulnerable to adversarial examples, which maliciously designed by adding small perturbations benign inputs imperceptible the human eye, leading false predictions target model. Compared character- and sentence-level textual attacks, word-level attack can generate higher-quality especially in a black-box setting. However, existing methods usually require huge number of queries successfully deceive model, is costly real...