28 June 2024 in IEEE Transactions on Mobile Computing
Recently, the continuous evolution of deep learning has opened up promising avenues to groundbreaking advancements in wireless sensing systems, which significantly enhance the practical applications of WiFi-based Human Activity Recognition (HAR) systems. However, despite these strides, such systems remain susceptible to adversarial attacks. This article unveils the vulnerability of existing WiFi-based HAR systems to common adversaries, revealing their insufficient robustness. While the intuitive approach is to employ adversarial training to fortify the models, our investigation exposes inherent deficiencies in the current approach. Specifically, we confirm that the strength of perturbations directly influences training outcomes. Moreover, even when confined within a specified perturbation radius, the perturbation strength exhibits variability within a prescribed range, potentially giving rise to “extreme” samples that could compromise training results. To address this challenge, we propose a two-stage Selective Adversarial Training (SAT) approach that integrates model confidence calibration and sample selection. Specifically, we start with calibrating the model and then selectively choose samples from all adversarial examples based on the calibrated confidence outputs that align with the desired criteria for adversarial training. This sample-wise perturbation intensity control effectively prevents the inclusion of inappropriate samples in training, a capability lacking in previous domain-wise perturbation control. Our experiments demonstrate that the proposed fine-grained training method, SAT, is both straightforward and effective in augmenting adversarial training results.
10 January 2024 in IEEE Transactions on Evolutionary Computation
AI-aided Financial Regulation (AIFR) is a practical and significant task, but current solutions have yet to be optimized with customized model designs. Given the privacy concerns surrounding financial data, we aim to employ Neural Architecture Search (NAS) to help non-expert end-users automatically design architectures. The genetic algorithm-based NAS stands out due to its relatively low hardware requirements and robust theoretical foundation. However, constrained by limited data, the model would undergo architecture search on a general regulatory dataset while being deployed on private one owned by each organization. The data distribution of the private dataset may vary from that of public datasets, giving rise to the challenge of data domain shift. To alleviate this problem, we propose a novel fitness evaluation method. When scoring the fitness, we take into account both the architecture’s validation accuracy and its potential for generalization by the metric of loss landscape. In addition, we improve the training paradigm for evaluation, utilizing a prototype-based training paradigm based on embedding distances for classification, allowing for rapid domain adaptation and improve performance on the distribution-shift data. We further introduce GA-TextCNN, a GA-based NAS framework specifically designed for text recognition, enhancing its suitability for text data within AIFR tasks. To demonstrate the effectiveness of our approach, we collect two related datasets and evaluate our method on it. The extensive experiments demonstrate that our method significantly improves baseline models and is effective in solving the AIFR problem.
28 May 2025 in IEEE Transactions on Artificial Intelligence
In recent years, deep learning-based Chinese font style transfer has garnered extensive research. This field not only opens new possibilities for artistic creation but also provides powerful tools for generating personalized digital content, especially in the thriving era of AIGC. The complexity and diversity of Chinese characters pose significant challenges for font style transfer, prompting numerous solutions proposed by the research community. In light of these rapid advancements, we aim to provide a comprehensive review of the latest developments in Chinese font style transfer. Specifically, we first outline traditional methods, particularly those that existed before the emergence of deep learning techniques, to establish a theoretical foundation. Subsequently, we delve into the current mainstream deep learning methods, including Autoencoder, Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs), and their applications in Chinese font style transfer. While these methods have shown remarkable performance in handling the shapes, structures, and artistic styles of Chinese characters, they still face inherent challenges and limitations. Therefore, we propose an innovative set of solutions aimed at overcoming these obstacles and improving conversion effectiveness and practicality. This is the first comprehensive study on font style transfer methods based on deep learning technology. We hope to provide new ideas and directions for future research and application of Chinese font style transfer. The undertaking of this work enriches theoretical research on style transfer and offers substantial support for practical applications, possessing profound significance and widespread impact.