Kexin Li

Kexin Li, Aastha Mehta, David Lie

To protect privacy in regulated domains such as healthcare and finance, model owners may allow only remote API access while keeping both the training data and model parameters private. However, model users performing inference on such remotely hosted models may be required to transmit potentially sensitive inputs, raising privacy concerns. In this work, we present LDPKiT, a framework for non-adversarial, privacy-preserving model distillation that leverages a user's private in-distribution data while bounding privacy leakage. LDPKiT introduces a novel superimposition technique that generates approximately in-distribution samples, enabling effective knowledge transfer under local differential privacy (LDP). Experiments on Fashion-MNIST, SVHN, and PathMNIST demonstrate that LDPKiT consistently improves utility while maintaining privacy, with benefits that become more pronounced at stronger noise levels. For example, on SVHN, LDPKiT achieves nearly the same inference accuracy at $ε=1.25$ as at $ε=2.0$, yielding stronger privacy guarantees with less than a 2\% accuracy reduction. We further conduct sensitivity analyses to examine the effect of dataset size on performance and provide a systematic analysis of latent space representations, offering intuitive and empirical insights into the accuracy gains of LDPKiT.

Yuxuan Wang, Yijun Liu, Fei Yu et al.

Despite the rapid development of Chinese vision-language models (VLMs), most existing Chinese vision-language (VL) datasets are constructed on Western-centric images from existing English VL datasets. The cultural bias in the images makes these datasets unsuitable for evaluating VLMs in Chinese culture. To remedy this issue, we present a new Chinese Vision- Language Understanding Evaluation (CVLUE) benchmark dataset, where the selection of object categories and images is entirely driven by Chinese native speakers, ensuring that the source images are representative of Chinese culture. The benchmark contains four distinct VL tasks ranging from image-text retrieval to visual question answering, visual grounding and visual dialogue. We present a detailed statistical analysis of CVLUE and provide a baseline performance analysis with several open-source multilingual VLMs on CVLUE and its English counterparts to reveal their performance gap between English and Chinese. Our in-depth category-level analysis reveals a lack of Chinese cultural knowledge in existing VLMs. We also find that fine-tuning on Chinese culture-related VL datasets effectively enhances VLMs' understanding of Chinese culture.

Kexin Li

Iteratively reweighted L1 (IRL1) algorithm is a common algorithm for solving sparse optimization problems with nonconvex and nonsmooth regularization. The development of its acceleration algorithm, often employing Nesterov acceleration, has sparked significant interest. Nevertheless, the convergence and complexity analysis of these acceleration algorithms consistently poses substantial challenges. Recently, Anderson acceleration has gained prominence owing to its exceptional performance for speeding up fixed-point iteration, with numerous recent studies applying it to gradient-based algorithms. Motivated by the powerful impact of Anderson acceleration, we propose an Anderson-accelerated IRL1 algorithm and establish its local linear convergence rate. We extend this convergence result, typically observed in smooth settings, to a nonsmooth scenario. Importantly, our theoretical results do not depend on the Kurdyka-Lojasiewicz condition, a necessary condition in existing Nesterov acceleration-based algorithms. Furthermore, to ensure global convergence, we introduce a globally convergent Anderson accelerated IRL1 algorithm by incorporating a classical nonmonotone line search condition. Experimental results indicate that our algorithm outperforms existing Nesterov acceleration-based algorithms.