Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Fei Peng has a total of 148 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Multimodal Resource-Exhaustion Attacks on Vision-Language Models via Joint Pixel-Prompt Optimization

    by Zhaoxiong Ni, Yatie Xiao, Chi-Man Pun, Fei Peng, Qingxiao Guan, Keke Tang

  • Flexural strength of FRP reinforced UHPC beams considering size effects

    by Xinlong Chen, Fei Peng, Weichen Xue

    Available studies have shown that the contribution of steel fibers to the flexural capacity of ultra-high-performance concrete (UHPC) beams is significant and inherently size-dependent. This study proposes a design-oriented method to determine the flexural strength of reinforced UHPC beams internally with fiber-reinforced polymer (FRP) bars. To this end, a validated analytical model was first developed for estimating the moment-curvature of FRP-reinforced UHPC sections. Subsequently, a parametric investigation was carried out to examine the effects of steel fiber inclusion in UHPC and section depth on flexural strength. The results showed that crack-bridging steel fibers significantly enhanced the flexural capacity of tension failure-controlled sections, whereas their contribution was insignificant in compression failure-controlled sections. Neglecting the steel fiber contribution may lead to incorrect predictions of the failure mode. A pronounced size effect was evident in the flexural strength as well as in the balanced reinforcement ratio. At an FRP reinforcement ratio of 1.0%, increasing the beam depth from 250 mm to 2000 mm resulted in an approximate 27.5% reduction in normalized flexural strength. Finally, a design-oriented flexural strength method that accounts for both steel fiber contribution and size effect was proposed. The accuracy of the proposed method was validated through comparison with available experimental data from flexural tests on 42 UHPC beams.

  • AdvDiffusion: Adversarial Patches Generation for Face Recognition With High Transferability in Physical Domain

    by Fei Peng, Yang Liu, Guohui Zhou, Min Long

    Face recognition models are vulnerable to spoofing of adversarial patches in the physical world. Attackers can enable face recognition models to make false identity judgments by simply pasting a sticker with a special pattern on the face. However, existing attacks lack the ability to transfer to black-box models, and the improvement of transferability is mainly focused on adversarial perturbations based on the p-norm. To further improve the attack performance and transferability, a highly transferable face recognition adversarial patches generation method named as AdvDiffusion is proposed. It first determines the region for adversarial patches generation based on facial gradient maps, and then an image is reconstructed to generate an adversarial patch by adding noise and denoising it with a pre-trained diffusion model. In the denoising, an adversarial loss is used to fine-tune the model and control the image to generate an adversarial patch with spoofing capability. Experiments and analysis show that the adversarial patches generated by the proposed mehtod have good adversarial attack capability on black-box face recognition models in both digital and physical domains, and also have better robustness under the changes of a complex physical environment compared with some state-of-the-art methods. It has great potential application for black-box attacks in the physical domain.

  • Proactive Defense for Physical-World 3D Adversarial Face Presentation Attacks

    by Le Qin, Min Long, Fei Peng, Shaowei Wang

    To defend against 3D adversarial face presentation attacks (PAs), this paper proposes an initiative defense scheme to proactively prevent face images from being misused by malicious attackers for crafting adversarial presentation attack instruments (PAIs). The proposed scheme is motivated by theoretically investigating the impostor attack principle behind 3D adversarial PAs, and its core idea is to interfere with the facial biometric fitting process of such attacks by injecting proactive defense components into face images beforehand. To learn the defense components, a hierarchical loss function is constructed based on attack objective disruption and fitting direction disruption, and it is able to well supervise the generation of the defense components and enables them to disturb the ultimate identity impersonation goals and the intermediate optimization orientations of 3D adversarial PAs. To analyze the performance of the proposed scheme, experiments are carried out with four different PAIs crafted by the state-of-the-art 3D adversarial PAs. The results indicate that it can achieve high defense success rates for the 3D adversarial PAIs based on eyes, eyes and nose, respirator, and natural style changes while preserving reasonable facial fidelity, and it exhibits generalization ability for unseen face recognition models, JPEG compression, resizing, and potential anti-defense approaches to a certain extent. Furthermore, compared with the existing proactive defense methods of face images, it obtains better performance when countering the 3D adversarial face PAs.


  • Large Capacity H.265/HEVC Video Steganography Based on Polygon Encoding and Improved Deep Learnable Similarity Network

    by Jiachen Xie, Xiang Zhang, Zhangjie Fu, Fei Peng, Fan Wang, Wenbin Huang, Daoyong Fu, Min Long

    In recent years, video steganography technique based on H.265/HEVC has received widespread attention. Typically, video steganography selects various syntax elements during the encoding process as carriers, and utilizing Prediction Unit (PU) as carrier is currently one of the most significant research directions. However, due to the limited number of PU types, such algorithms often suffer from insufficient capacity and visual quality. To alleviate the aforementioned issues, this paper proposes an H.265/HEVC video steganography algorithm that utilizes polygon encoding and Improved Deep Learnable Similarity Network Filter (IDLSNF). Firstly, we design a new polygon encoding rule, which maps different integers into several polygons. Secondly, we propose a novel steganography method based on polygon encoding and PU partition mode. This method selects the PUs of $8\times 8$ and $16\times 16$ coding unit in P-frames as carriers and hides the secret message by modifying the partition mode of two adjacent PUs. Due to the ability of polygon encoding to represent more information within a small range, it increases capacity with low steganographic distortion. Thirdly, we further propose a filter by improving DLSN, which enhances the visual quality of the entire stego video by processing I-frames. Extensive experimental results show that the video steganography algorithm proposed in this paper achieves higher capacity and superior visual quality compared to current State-of-the-Art methods. Meanwhile, our algorithm can also obtain good BRI and anti-steganalysis performance. This method has promising application prospects in the field of video covert communication.


  • Harnessing Transferable Adversarial Examples via Multilayer Attention-Guided Spatial Transformations

    by Pengfei Dong, Yatie Xiao, Chi-Man Pun, Fei Peng, Kongyang Chen, Qingxian Guan, Siyuan Chen, Xiangyu Ye, Zhenbang Liu

    Transfer-based adversarial attacks are key for evaluating the robustness of deep neural networks (DNNs) in black-box settings, yet their effectiveness is often constrained by limited cross-model transferability. Existing feature-level approaches typically rely on single-layer attention guidance or static perturbation patterns, which restrict adaptability across diverse architectures. In this work, we introduce a unified adversarial framework, named Multi-layer Attention-guided Spatial Transformations (MAT), to exploit class-discriminative cues from multiple feature layers to craft highly transferable adversarial examples. MAT integrates Multi-layer Attention Fusion (MAF) to capture complementary low-level and high-level semantics from multiple intermediate layers, Attention-guided Augmentation (AGA) to selectively perturb non-critical regions while preserving semantic integrity, and Spatial Random Transformation (SRT) to introduce stochastic spatial augmentations to diversify patterns during optimization. Unlike prior methods that use static or layer-specific attention, MAT dynamically adapts feature guidance to the architecture and task, which enhances generalization. We evaluate MAT against eleven state-of-the-art (SOTA) transfer-based attacks across nine CNN-based and Transformer-based architectures on ImageNet. Comprehensive experiments demonstrate that MAT consistently outperforms eleven state-of-the-art transfer-based attacks in both white-box and black-box settings, including against adversarially trained and input preprocessing-based defensive models, while maintaining higher semantic similarity to the original inputs. It highlights the superior adversarial robustness and excellent adaptability of MAT in adversarial machine learning. Our code is available at https://github.com/dislab-gzhu/MAT.

  • AMG: Attention Triple-Mask-Guided Screen-Shooting Resilient Watermarking

    by Fei Peng, Jianhang Xu, Shenghui Zhu, Min Long

    To address the growing demand for digital image security, this paper introduces AMG, a U-Net-based robust watermarking framework designed for screen-shooting resilience and watermark imperceptibility. The framework fuses channel and spatial attention to adaptively enhance key region features during encoding and decoding, achieving precise watermark embedding with visual fidelity. Guided by a Triple Mask of depth, edge, and gradient components, a specialized loss function strategically embeds the watermark into structurally-rich yet visually non-salient regions. Experiments show AMG surpasses state-of-the-art methods in extraction accuracy, robustness, and imperceptibility, offering a systematic solution to balance watermark security and visual quality.