Traditional Chinese Medicine (TCM), as an effective alternative medicine, has been receiving increasing attention. In recent years, the rapid development of large language models (LLMs) tailored for TCM has highlighted the urgent need for an objective and comprehensive evaluation framework to assess their performance on real-world tasks. However, existing evaluation datasets are limited in scope and primarily text-based, lacking a unified and standardized multimodal question-answering (QA) benchmark. To address this issue, we introduce TCM-Ladder, the first comprehensive multimodal QA dataset specifically designed for evaluating large TCM language models. The dataset covers multiple core disciplines of TCM, including fundamental theory, diagnostics, herbal formulas, internal medicine, surgery, pharmacognosy, and pediatrics. In addition to textual content, TCM-Ladder incorporates various modalities such as images and videos. The dataset was constructed using a combination of automated and manual filtering processes and comprises over 52,000 questions. These questions include single-choice, multiple-choice, fill-in-the-blank, diagnostic dialogue, and visual comprehension tasks. We trained a reasoning model on TCM-Ladder and conducted comparative experiments against nine state-of-the-art general domain and five leading TCM-specific LLMs to evaluate their performance on the dataset. Moreover, we propose Ladder-Score, an evaluation method specifically designed for TCM question answering that effectively assesses answer quality in terms of terminology usage and semantic expression. To the best of our knowledge, this is the first work to systematically evaluate mainstream general domain and TCM-specific LLMs on a unified multimodal benchmark. The datasets and leaderboard are publicly available at https://tcmladder.com and will be continuously updated. The source code is available at https://github.com/orangeshushu/TCM-Ladder.

Large volumes of images are being exponentially generated today, which poses high demands on the services of storage, processing, and management. To handle the explosive image growth, a natural choice nowadays is cloud computing. However, coming with the cloud-based image services is acute data privacy concerns, which has to be well addressed. In this paper, we present a secure cloud-based image service framework, which allows privacy-preserving and effective image denoising on the cloud side to produce high-quality image content, a key for assuring the quality of various image-centric applications. We resort to state-of-the-art image denoising techniques based on deep neural networks (DNNs), and show how to uniquely bridge cryptographic techniques (like lightweight secret sharing and garbled circuits) and image denoising in depth to support privacy-preserving DNN based image denoising services on the cloud. By design, the image content and the DNN model are all kept private along the whole cloud-based service flow. Our extensive empirical evaluation shows that our security design is able to achieve denoising quality comparable to that in plaintext, with high cost efficiency on the local side and practically affordable cost on the cloud side.

In crowd sensing, truth discovery (TD) refers to finding reliable information from noisy/biased data collected from different providers. To protect providers' data while enabling truth distillation, privacy-preserving truth discovery (PPTD) has received wide attention recently. However, all existing approaches require iterative interaction between server(s) and individual providers, which inevitably demand all providers to be always online. Otherwise, the protocol would fail or expose extra provider information. In this paper, we design and implement the first non-interactive PPTD system that completely removes the online requirement with strong privacy guarantees. Our framework follows the same two-server model from the best-known prior solution, and leverages Yao's Garbled Circuit (GC). Yet, we devise non-trivial speedup techniques for TD-optimized implementation. Firstly, we identify reusable computations in TD to accelerate the circuit generation. Secondly, we securely evaluate the burdensome non-linear functions in TD via customized approximation with accuracy and improved efficiency. Thirdly, we reduce the online execution time by bridging together latest advancements of component-based GC and various computations needed in TD. Unlike prior arts, our framework does not reveal any intermediate results, and further supports “late-join” providers without protocol suspension/restart. The practical performance of our proof-of-concept implementation is verified through extensive evaluations.