Bingbing Song
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Bingbing Song has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Wei Zhou
School of Software, Yunnan University, Kunming, China
8 shared publications - Jinhong Zhang
Yunnan University
3 shared publications - Yunyun Dong
Yunnan University
3 shared publications - Song Gao
Yunnan University
2 shared publications - Chao Yi
National Pilot School of Software Yunnan University 650504 Kunming Yunnan China
1 shared publication - Chao Yi1 shared publication
Latest Publications
- Open Access
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Personalized visual privacy localization with user-defined concepts
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.
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPersonalized visual privacy localization with user-defined concepts
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.
- 21 January 2026
- Published by Elsevier in Neural Networks
- Vol. 198, 108626
RWP: a robust watermarking plugin for attribution and protection in stable diffusion models
- 21 January 2026
- Published by Elsevier in Neural Networks
- Vol. 198, 108626
RWP: a robust watermarking plugin for attribution and protection in stable diffusion models
- 16 November 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 302
CRFD: A novel face privacy preservation via fine-grained controllable and reversible de-identification
- 16 November 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 302
CRFD: A novel face privacy preservation via fine-grained controllable and reversible de-identification
- 7 November 2024
- Published by Springer Nature
- p. 160-173
Generative Steganography Based on Dual-Branch Flow
- 7 November 2024
- Published by Springer Nature
- p. 160-173
Generative Steganography Based on Dual-Branch Flow
- Top Cited Papers
- 5 June 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 254
A survey on Deep-Learning-based image steganography
- 5 June 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 254
Top Cited PapersA survey on Deep-Learning-based image steganography
- 4 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 8 (3) , 2518-2528
Exploiting Type I Adversarial Examples to Hide Data Information: A New Privacy-Preserving Approach
Deep neural networks (DNNs) are sensitive to adversarial examples which are generated by corrupting benign examples with imperceptible perturbations, or have significant changes but can still achieve original prediction results. The latter case is termed as the Type I adversarial example which, however, has limited attention in the literature. In this paper, we introduce two methods, termed HRG and GAG, to generate Type I adversarial examples and attempt to apply them to the privacy-preserving Machine Learning as a Service (MLaaS). Existing methods for the privacy-preserving MLaaS are mostly based on cryptographic techniques, which often incur additional communication and computation overhead, while using Type I adversarial examples to hide users' privacy data is a brand-new exploration. Specifically, HRG utilizes the high-level representations of DNNs to guide generators, and GAG leverages the generative adversarial network to transform original images. Our solution does not involve any model modifications and allows DNNs to run directly on transformed data, thus arousing no additional communication and computation overhead. Extensive experiments on MNIST, CIFAR-10, and ImageNet show that HRG can perfectly hide images into noise and achieve similar accuracy as the original accuracy, and GAG can generate natural images that are completely different from the original images with a small loss of accuracy.
- 4 March 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 8 (3) , 2518-2528
Exploiting Type I Adversarial Examples to Hide Data Information: A New Privacy-Preserving Approach
Deep neural networks (DNNs) are sensitive to adversarial examples which are generated by corrupting benign examples with imperceptible perturbations, or have significant changes but can still achieve original prediction results. The latter case is termed as the Type I adversarial example which, however, has limited attention in the literature. In this paper, we introduce two methods, termed HRG and GAG, to generate Type I adversarial examples and attempt to apply them to the privacy-preserving Machine Learning as a Service (MLaaS). Existing methods for the privacy-preserving MLaaS are mostly based on cryptographic techniques, which often incur additional communication and computation overhead, while using Type I adversarial examples to hide users' privacy data is a brand-new exploration. Specifically, HRG utilizes the high-level representations of DNNs to guide generators, and GAG leverages the generative adversarial network to transform original images. Our solution does not involve any model modifications and allows DNNs to run directly on transformed data, thus arousing no additional communication and computation overhead. Extensive experiments on MNIST, CIFAR-10, and ImageNet show that HRG can perfectly hide images into noise and achieve similar accuracy as the original accuracy, and GAG can generate natural images that are completely different from the original images with a small loss of accuracy.
- 15 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Services Computing
- Vol. 16 (5) , 3723-3735
Securing Deep Learning as a Service Against Adaptive High Frequency Attacks With MMCAT
Most cloud providers offer Deep Learning as a Service (DLaaS) for different business, science and engineering domains. However, it is known that deep neural networks (DNNs) are vulnerable to adversarial examples, which can cause well-trained DNN models to misbehave by injecting human-imperceptible perturbations to the query input data. Securing deep learning as a service becomes a critical challenge in mitigating such adversarial input perturbations, and enhancing the robustness of DNNs. In this paper, we report two important facts: First, most adversarial perturbations are high frequency signals or are added to high frequency signals. Second, due to Frequency Principle that neural networks overly pay attention to fit the low frequency signals during training, the models could be easily misled by the high frequency signals of adversarial examples. These facts consequently contribute to the vulnerability of DNNs service in the Cloud. We conjecture that the more robust the neural networks are in learning from high frequency signals, the more resilient these neural networks are against adversarial perturbed examples. We propose a novel method for generating high-frequency-enhanced adversarial examples, which is achieved by a high-pass filter in the frequency domain via Fourier Transform. This method enhances the learning ability for high frequency signals and ameliorates to over-fit useless low frequency signals. In order to improve the robustness of DNNs service under such signal frequency attacks, we propose a multi-modal collaborative adversarial training framework, named as MMCAT, which uses the multi-modal information of the input images for cross-modal collaborative training, delivering excellent extension for effectively learning of multi-modal image information. Extensive experiments show that under strong adaptive frequency attacks, the DNNs service trained with the proposed MMCAT method achieve superior performance and robustness over the state-of-the-art adversarial training approaches.
- 15 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Services Computing
- Vol. 16 (5) , 3723-3735
Securing Deep Learning as a Service Against Adaptive High Frequency Attacks With MMCAT
Most cloud providers offer Deep Learning as a Service (DLaaS) for different business, science and engineering domains. However, it is known that deep neural networks (DNNs) are vulnerable to adversarial examples, which can cause well-trained DNN models to misbehave by injecting human-imperceptible perturbations to the query input data. Securing deep learning as a service becomes a critical challenge in mitigating such adversarial input perturbations, and enhancing the robustness of DNNs. In this paper, we report two important facts: First, most adversarial perturbations are high frequency signals or are added to high frequency signals. Second, due to Frequency Principle that neural networks overly pay attention to fit the low frequency signals during training, the models could be easily misled by the high frequency signals of adversarial examples. These facts consequently contribute to the vulnerability of DNNs service in the Cloud. We conjecture that the more robust the neural networks are in learning from high frequency signals, the more resilient these neural networks are against adversarial perturbed examples. We propose a novel method for generating high-frequency-enhanced adversarial examples, which is achieved by a high-pass filter in the frequency domain via Fourier Transform. This method enhances the learning ability for high frequency signals and ameliorates to over-fit useless low frequency signals. In order to improve the robustness of DNNs service under such signal frequency attacks, we propose a multi-modal collaborative adversarial training framework, named as MMCAT, which uses the multi-modal information of the input images for cross-modal collaborative training, delivering excellent extension for effectively learning of multi-modal image information. Extensive experiments show that under strong adaptive frequency attacks, the DNNs service trained with the proposed MMCAT method achieve superior performance and robustness over the state-of-the-art adversarial training approaches.
- Open Access
- 10 October 2022
- Published by MDPI AG in Mathematics
- Vol. 10 (19) , 3716
Robustness Learning via Inference-Softmax Cross Entropy in Misaligned Distribution of Image
Adversarial examples easily mislead vision systems based on deep neural networks (DNNs) trained with softmax cross entropy (SCE) loss. The vulnerability of DNN comes from the fact that SCE drives DNNs to fit on the training examples, whereas the resultant feature distributions between the training and adversarial examples are unfortunately misaligned. Several state-of-the-art methods start from improving the inter-class separability of training examples by modifying loss functions, where we argue that the adversarial examples are ignored, thus resulting in a limited robustness to adversarial attacks. In this paper, we exploited the inference region, which inspired us to apply margin-like inference information to SCE, resulting in a novel inference-softmax cross entropy (I-SCE) loss, which is intuitively appealing and interpretable. The inference information guarantees that it is difficult for neural networks to cross the decision boundary under an adversarial attack, and guarantees both the inter-class separability and the improved generalization to adversarial examples, which was further demonstrated and proved under the min-max framework. Extensive experiments show that the DNN models trained with the proposed I-SCE loss achieve a superior performance and robustness over the state-of-the-arts under different prevalent adversarial attacks; for example, the accuracy of I-SCE is higher than SCE under the attack on the MNIST dataset. These experiments also show that the inference region can effectively solve the misaligned distribution.- 10 October 2022
- Published by MDPI AG in Mathematics
- Vol. 10 (19) , 3716
Open AccessRobustness Learning via Inference-Softmax Cross Entropy in Misaligned Distribution of Image
Adversarial examples easily mislead vision systems based on deep neural networks (DNNs) trained with softmax cross entropy (SCE) loss. The vulnerability of DNN comes from the fact that SCE drives DNNs to fit on the training examples, whereas the resultant feature distributions between the training and adversarial examples are unfortunately misaligned. Several state-of-the-art methods start from improving the inter-class separability of training examples by modifying loss functions, where we argue that the adversarial examples are ignored, thus resulting in a limited robustness to adversarial attacks. In this paper, we exploited the inference region, which inspired us to apply margin-like inference information to SCE, resulting in a novel inference-softmax cross entropy (I-SCE) loss, which is intuitively appealing and interpretable. The inference information guarantees that it is difficult for neural networks to cross the decision boundary under an adversarial attack, and guarantees both the inter-class separability and the improved generalization to adversarial examples, which was further demonstrated and proved under the min-max framework. Extensive experiments show that the DNN models trained with the proposed I-SCE loss achieve a superior performance and robustness over the state-of-the-arts under different prevalent adversarial attacks; for example, the accuracy of I-SCE is higher than SCE under the attack on the MNIST dataset. These experiments also show that the inference region can effectively solve the misaligned distribution.