JZJinhong Zhang

Jinhong Zhang

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Jinhong Zhang has a total of 25 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Personalized visual privacy localization with user-defined concepts

    by Junyu Hou, Jinhong Zhang, W. Shiung Liew, Wei Zhou, Bingbing Song

    Personalized visual privacy protection is crucial due to the subjective nature of privacy. However, existing methods typically rely on fixed, predefined categories, which struggle to generalize to open-ended, user-defined concepts. To overcome this limitation, we propose a novel task: Open-Ended User-Defined Visual Privacy Localization, which aims to localize arbitrary privacy concepts based on user-provided definitions without task-specific or user-specific fine-tuning. While existing datasets are mostly limited to image-level privacy classification or confined to fixed, predefined attributes, we construct the Concept-Level Visual Privacy (CLVP) dataset to facilitate research in this domain. Explicitly designed for open-ended, user-defined privacy concepts, CLVP provides high-quality pixel-level annotations for both private and non-private images, comprehensively covering both object-level and scene-level privacy expressions.We further propose VPGuard, a framework that effectively decouples privacy reasoning from visual localization. Unlike conventional approaches that rely on direct segmentation, VPGuard employs a Definition-Aware Structured Prompt (DASP) strategy to guide a Multimodal Large Language Model (MLLM) in verifying privacy presence and generating definition-aligned referring expressions for a foundation segmentation model. Extensive experiments on the CLVP dataset and evaluations on established privacy dataset demonstrate that VPGuard significantly outperforms generic text-promptable segmentation approaches adapted for this task, offering a flexible solution for personalized privacy protection.

  • Open Access

    Source-Free Latent Watermarking for Robust and Scalable Image Traceability

    by Jinhong Zhang, Zongxin Liu, Wenqing Hong, Lizhen Liu, Wei Zhou, Bingbing Song

    Invisible watermarking supports ownership verification and leakage tracing in security-sensitive image distribution. Existing methods often rely on real images for training, encode user identities within limited payload bits, and remain vulnerable to diffusion-based semantic editing. These limitations are restrictive when protected images cannot be exposed and watermark evidence must survive both conventional and generative transformations. We propose a source-free latent watermarking framework for robust and scalable real-image traceability. The proposed framework learns watermark embedding and extraction from a diffusion-generated source distribution and operates in the frozen Stable Diffusion latent space, decoupling watermark learning from protected images while improving transfer to unseen domains. A combinatorial identity mapping assigns each user a reversible latent transformation, expanding the identity space without increasing the message length. The watermarking model is optimized with latent and image consistency losses together with clean and attack-aware decoding objectives.Experiments on natural, generated, artistic, and medical image domains achieve $98.80\%$ average adversarial bit accuracy rate and $0.9964$ average adversarial true positive rate. The watermark remains robust under diffusion-based image-to-image regeneration and inpainting, while supporting closed-set identification over candidate pools of up to $10^6$ users. These results demonstrate a practical trade-off among source-free training, robustness, visual fidelity, and scalable traceability.

  • Open Access

    CISC: Cognitively Imperceptible Generative Image Steganography with Security and Controllability

    by Yaxin Zeng, Jinhong Zhang, Wei Zhou, Ping Wei, Yunyun Dong

    Steganography aims to conceal secret information within ordinary media, providing an alternative to traditional encryption methods that may attract the attention of eavesdroppers. In recent years, cross-modal and generative steganography have leveraged multimodal data and diffusion models to enhance imperceptibility. However, existing approaches often provide limited controllability over generated content and may introduce security concerns when images and texts are transmitted simultaneously, as the accompanying text can potentially be exploited to facilitate message recovery.To address these challenges, we propose \textbf{CISC}: \textbf{C}ognitively \textbf{I}mperceptible Generative Image Steganography with \textbf{S}ecurity and \textbf{C}ontrollability. CISC incorporates two main designs: a mapping-sampling strategy and a dual-layer text transformation mechanism. The mapping-sampling strategy allows stego image generation to be guided by user-specified prompts, supporting flexible manipulation of visual content. Meanwhile, the dual-layer text transformation mechanism aims to reduce prompt-level information exposure during multimodal transmission.Experimental results show that CISC can generate controllable stego images while maintaining visual inconspicuousness, and it provides competitive performance in robustness and security-related evaluations, suggesting a practical pathway toward cognitively imperceptible and secure multimodal steganography.

  • RWP: a robust watermarking plugin for attribution and protection in stable diffusion models

    by Zongxin Liu, Jinhong Zhang, Yunyun Dong, Bingbing Song, Wei Zhou


  • Model Inversion Attacks on Homogeneous and Heterogeneous Graph Neural Networks

    by Renyang Liu, Wei Zhou, Jinhong Zhang, Xiaoyuan Liu, Peiyuan Si, Haoran Li


  • AFLOW: Developing Adversarial Examples Under Extremely Noise-Limited Settings

    by Renyang Liu, Jinhong Zhang, Haoran Li, Jin Zhang, Yuanyu Wang, Wei Zhou


  • Open Access

    DeepLSH: A Deep Learning-based Locality Sensitive Hash Method

    by Renyang Liu, Qiwei Dou, Haoran Li, Jinhong Zhang, Junhui Liu

    Approximate nearest neighbor (ANN) search plays a vital role in the database system, data retrieval, and any scenario where similar elements need to be queried in a fast manner. The traditional ANN tasks are typically completed by hash-style techniques, such as the locality-sensitive-hash (LSH). However, while the amount of data is consistently rising in modern real applications, an ANN system is required to scale to the larger data size as well as the increasing data dimension. Considering that the hash functions struggle with precision and efficiency on such a dataset, the hash-based ANN systems are becoming unsuitable for big data environments. By contrast, the learning-based methods provide a promising way for handling high-dimensional data, which also reach a good balance between precision and efficiency in ANN tasks. Starting from this, we advocate that a pleasing way to improve the ANN system is to leverage both the deep learning and the distribution information. We propose a practical deep learning-based hash framework, which is termed as deep learning-based locality-sensitive-hash (DeepLSH). This framework optimizes the binary hash code by building up an encoder-decoder deep neural network which can be trained in an end-to-end manner. Experimental results show that our method satisfies the generalized LSH and outperforms the existing hashing algorithms on challenging datasets.