Keynote Speaker

Title: IoT-Based Electronic Health Records Protection Using Compressed Images as Carriers

Abstract: Since the COVID-19 outbreak, there’s been a growing need for contactless healthcare to meet medical diagnosis demands. Electronic health systems using the Internet of Things (IoT) are rapidly advancing, transmitting significant amounts of private medical data online. In telemedicine, where patients are diagnosed remotely, sensitive information such as patient records may be embedded into medical images for security purposes. Due to the large file sizes of medical images produced by equipment, compression is essential for fast transmission.
To safeguard medical images in telemedicine and address bandwidth limitations, people utilize data hiding techniques and absolute moment block truncation coding (AMBTC) compression to introduce an IoT-driven electronic health protection mechanism. In this talk， I will introduce a mechanism which employs diverse methods to compress and embed data across various image blocks. Additionally, it offers a flexible adaptation to meet different application requirements concerning embedding capacity, visual quality, and file size by adjusting thresholds and variants. Compared to alternative methods, this approach delivers superior payload capacity and efficiency while preserving visual fidelity.

Biography

Prof. Chin-Chen Chang
IEEE Fellow
Feng Chia University, China

Title: Vision-Language Models: Pre-Training, Fine-Tuning and Trustworthiness

Abstract: Vision-language models (VLMs) are now widely used in various vision-language modeling tasks. However, there are still a number of challenges. First, cross-modal masked language modeling is often used to learn the vision-language associations. However, existing masking strategies are insufficient in that the masked tokens can sometimes be simply recovered with only the language information, ignoring the visual inputs. Second, during fine-tuning, multiple models with various hyperparameter configurations are often created, but typically only one of these models is actually utilized in the downstream task. Third, vision-language models are more vulnerable to jailbreak attacks than their LM predecessors.

To address the first issue, we use a masking strategy based on the saliencies of language tokens to the image. For the second issue, we consider the learned soup, which combines all fine-tuned models with learned weighting coefficients. While this can significantly enhance performance, it is also computationally expensive. We propose to mitigate this by formulating the learned soup as a computationally-efficient hyperplane optimization problem and employing block coordinate gradient descent to learn the mixing coefficients. Finally, to construct robust VLMs, we propose a training-free protecting approach that exploits the inherent safety awareness of LLMs, and generates safer responses via adaptively transforming unsafe images into texts to activate the intrinsic safety mechanism of pre-aligned LLMs in VLMs.
 

Biography

Prof. James Kwok
IEEE Fellow
Hong Kong University of Science and Technology, Hongkong

Title: Information Theoretic Learning

Abstract: Information theory has attracted increasing attention in the fields of machine learning and signal processing in recent years. Novel information theoretic approaches have been proposed for different learning problems, such as supervised learning with the minimum error entropy (MEE) criterion, and representation learning with the information bottleneck (IB) principle. This talk introduces the basic principles and paradigms of information theoretic learning (ITL), and discusses the applications in brain inspired computing, brain computer interfaces and brain disease diagnosis.

Biography

Prof. Badong Chen
Institute of Artificial Intelligence and Robotics, XJTU
College of Artificial Intelligence, XJTU
School of Electronic and Inform. Engineering, XJTU