Oct 2026· International Conference on Optical Communication and Optical Information Processing· Vol 14501, pp. 145010V - 145010V-8· 0 citations· 17 references
Engineering
TL;DR
Ortho Fusion (OF), an imperceptible 3D point cloud adversarial attack framework based on frequency-domain features and hierarchical gradient orthogonalization, is proposed, which maintains a high attack success rate while significantly improving the imperceptibility and geometric quality of adversarial examples.
Abstract
The vulnerability of 3D point cloud deep learning models to adversarial perturbations has attracted increasing attention in recent years. However, existing point cloud adversarial attacks often suffer from severe multi-objective gradient conflicts when simultaneously optimizing attack success and geometric imperceptibility. To address this issue, we propose Ortho Fusion (OF), an imperceptible 3D point cloud adversarial attack framework based on frequency-domain features and hierarchical gradient orthogonalization. The proposed method first decomposes the input point cloud into low-frequency and high-frequency components and restricts perturbation optimization to the low-frequency manifold to preserve global geometric consistency. It then employs a hierarchical asymmetric PCGrad mechanism to alleviate conflicts among full-frequency adversarial objectives, low-frequency enhancement objectives, and geometric constraints through sequential orthogonal projections. Experimental results demonstrate that OF maintains a high attack success rate while significantly improving the imperceptibility and geometric quality of adversarial examples.
A novel adversarial attack resilient network, called Frequency-Aware and Dynamic Curve Network for Adversarially Robust Infrared and Visible Image Fusion (FDCNet), and an Extrema-guided Dynamic Tone Curve Module (EDTC) at the back-end as a post-compensation mechanism to restore pixel distributions.
Peng-Cheng Gao, Sheng-Yue Huang· Journal of King Saud Univers...· 0 citations
The Enhanced Deformation Attack (EDA), which estimates attack gradients over stochastically transformed views of the current adversarial image, achieves the highest mean attack success rate (ASR) on ViT targets for all four CNN sources, outperforming the strongest source-specific baseline.
Understanding adversarial attacks is crucial for the secure deployment of visual recognition systems. While ensemble attacks combine the strengths of transfer-based and query-based methods to generate highly transferable adversarial examples, their practicality is often limited in real-world scenarios where the number...
Zhun Zhang, Jian Wang, Shi-Ze Guo et al.· IEEE Transactions on Informa...· 0 citations
Transferable adversarial attacks provide an important means of evaluating the black-box security of remote sensing scene classification models. However, the existing spatial input transformations commonly rely on predefined block shapes and limited partition granularities, providing insufficient coverage of the diverse...
Rui Zhang, Jie Wang, Wen-Jun Hu et al.· IEEE Geoscience and Remote S...· 0 citations
The Information Blackhole principle is proposed, which introduces Gaussian distribution constraints to disentangle latent representations and block interfering information, and Adaptive Gaussian Matching (AGM), which explicitly regularizes the latent distribution of poisoned samples.
Zhi-Fei Yang, Xiu-Ping Liu, Kuo-Feng Gao et al.· 0 citations
Uncrewed aerial vehicles increasingly rely on on-board deep object detectors for applications such as surveillance, delivery, agriculture, and traffic monitoring. The vulnerability of these detectors to adversarial patches has therefore become a practical safety concern. Research on adversarial patch and camouflage att...
Sandesh Shrestha, K. T. Yasas Mahima, Asanka G. Perera· Italian National Conference...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.