HUANG: A Robust Diffusion Model-based Targeted Adversarial Attack Against Deep Hashing Retrieval
Authors: Chihan Huang, Xiaobo Shen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superiority of HUANG across different datasets, achieving state-of-the-art performance in black-box targeted attacks. Additionally, the dynamic interplay between denoising and adding adversarial perturbations in adversarial denoising endows HUANG with exceptional robustness and transferability. Comprehensive experiments reveal that HUANG greatly outperforms prior models, setting a new benchmark for targeted black-box attacks. Evaluation against adversarial defense and other assessments validate the model s superior capabilities in terms of transferability and robustness. Table 1: The targeted attack performance comparison between HUANG and other advanced attack methods. The evaluation metric is t-MAP. Figure 2: Visual comparison of benign images and adversarial images generated by HUANG on three datasets. Figure 3: Comparison of t-MAP of HUANG and other SOTA algorithms after three kinds of defense methods. Table 2: t-MAP(%) and perceptibility ( 10 2) of different attack methods for the hashing model with 32-bits code length. Table 3: Transferability comparison result of HUANG and SAAT on DPSH. Figure 4: t-MAP and image quality of HUANG on three datasets when varying timestep T. Table 4: Ablation results on different module components on different datasets, here w/o means without. |
| Researcher Affiliation | Academia | Nanjing University of Science and Technology, Nanjing, China EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Adversarial denoising |
| Open Source Code | No | The paper does not contain any explicit statements about code release or links to a source code repository. |
| Open Datasets | Yes | We conduct experiments on three datasets: FLICKR-25K (Huiskes and Lew 2008), NUS-WIDE (Chua et al. 2009), and MS-COCO (Lin et al. 2014). |
| Dataset Splits | Yes | The FLICKR-25K dataset comprises 25,000 images labeled across 38 categories, from which we have selected 1,700 images as our query set, 5,000 images for the training set, and the remaining images to form the database. The NUS-WIDE dataset consists of 193,734 images labeled across 21 categories, with 2,100 images chosen as the query set, 10,000 for training, and the remaining images serving as the database. The MS-COCO dataset contains 123,287 images labeled across 80 categories, with 5,000 images designated as the query set, 10,000 for training, and the remaining images serving as the database. |
| Hardware Specification | Yes | The experimental setup is anchored by a hardware environment running Windows 11 and powered by a Ge Force RTX 4080 GPU. |
| Software Dependencies | Yes | The software environment utilizes Python 3.8 and the Py Torch 2.0.0 + cu118. |
| Experiment Setup | Yes | All images are resized to 224 224. The Adam optimizer is employed, with hyperparameters set to β1 = 0.9 and β2 = 0.999. The perturbation threshold is established at 8/255. |