QYQiwei Yu

Qiwei Yu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Qiwei Yu has a total of 10 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    BlockKD3A: A Web3-Native Framework for Decentralized Collaborative AI and User-Generated Model Assets

    by Ziang Zhou, Hongjian Shi, Ruhui Ma, Yuhan Qiu, Qiwei Yu, Shiquan Wang, Zhongle Qu, Haibing Guan

    User-generated model assets in Web3 demand verifiable provenance, clear attribution, and dependable coordination across organizational boundaries without centralizing data. However, real-world deployments still face persistent challenges. Even algorithmically robust methods like KD3A often lack auditable coordination, failure resilience, and reproducible artifact lineage. These limitations make training brittle and results difficult to audit amidst node churn, adversarial behavior, and system heterogeneity. We introduce BlockKD3A, a hybrid on-chain/off-chain framework that operationalizes KD3A with three key system guarantees: (i) auditable coordination, achieved via smart contracts that record training state, model content identifiers (CIDs), and Consensus Focus (CF)-based attribution; (ii) reliability, through a transaction-safe client and failure-aware orchestration that applies nonnegative CF clipping, supports zero-CF fallback, and avoids blocking on stragglers; and (iii) reproducibility, enabled by content-addressable packaging and unified telemetry that couples machine learning metrics with blockchain events. In 9 end-to-end deployments (40 epochs per target), BlockKD3A delivered near-centralized performance, 96.74% macro-average accuracy on Office-Caltech10, and 89.40% on DigitFive—substantially outperforming representative federated baselines. Coordination remained efficient and predictable, with only 13 blockchain transactions across all 9 deployments, bounded gas costs per write (125,972–298,405), and stable epoch timings even under partial participation. By integrating KD3A’s algorithm-native attribution with on-chain provenance and robust execution, BlockKD3A closes the gap between decentralized learning algorithms and reliable system deployment, providing auditability, predictable cost envelopes, and transferable engineering patterns for verifiable user-generated model assets.


  • MPCPFL: Prototype-Based Multimodal Personalized Federated Learning With Missing Modalities in Medical Domain

    by Qiwei Yu, Hongjian Shi, Zhuolong Jiang, Yuhan Qiu, Ruhui Ma, Shiliang Liao, Xing Zhou, Haibing Guan

    With the global aging population on the rise, the amount of disease data related to the elderly is also increasing. Effectively protecting personal data when analyzing and processing disease data has become crucial. Federated learning (FL), as a proven method, is widely used. In the multimodal FL for disease diagnosis, the intricate challenge of modality missing and heterogeneous data distribution across healthcare clients severely degrades model performance and generalization. Existing frameworks struggle to address these issues in scenarios with complex modality missing and data heterogeneity, particularly in privacy-sensitive medical applications. In this article, we introduce a novel framework, modal prototype and contrast for personalized federated learning, which enhances prototype-based multimodal federated learning with personalized local aggregation and enables the adaptive fusion of global prototypes and local client models during training. By leveraging the modal prototype to compensate for missing modalities and modal contrastive training to align client representations with shared prototypes, the framework mitigates global model degradation caused by modality absence. In addition, we adapt the SLAKE medical multimodal dataset into a VQAv2-style classification task, simulating realistic disease diagnosis scenarios with varying modality missing rates (10 $\sim$ 90%). Experimental results demonstrate that our approach achieves state-of-the-art performance. Compared to the baselines, our method improved inference accuracy by up to 13.7% when 90% of the modality was missing during training.