Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Vanish into Thin Air: Cross-prompt Universal Adversarial Attacks for SAM2
Authors: Ziqi Zhou, Yifan Hu, Yufei Song, Zijing Li, Shengshan Hu, Leo Yu Zhang, Dezhong Yao, Long Zheng, Hai Jin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on six datasets across two segmentation tasks demonstrate the effectiveness of the proposed method for SAM2. The comparative results show that UAP-SAM2 significantly outperforms state-of-the-art (SOTA) attacks by a large margin. |
| Researcher Affiliation | Academia | 1 National Engineering Research Center for Big Data Technology and System 2 Services Computing Technology and System Lab 3 Cluster and Grid Computing Lab 4 Hubei Engineering Research Center on Big Data Security 5 Hubei Key Laboratory of Distributed System Security School of Computer Science and Technology, Huazhong University of Science and Technology School of Cyber Science and Engineering, Huazhong University of Science and Technology School of Software Engineering, Huazhong University of Science and Technology School of Information and Communication Technology, Griffith University EMAIL EMAIL |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are present in the paper. The methodology is described through text and mathematical formulations. |
| Open Source Code | Yes | Our codes are available at: https://github.com/CGCL-codes/UAP-SAM2. |
| Open Datasets | Yes | We evaluate our attack on there public video segmentation datasets: You Tube VOS2018 (You Tube) [8], DAVIS 2017 (DAVIS) [26], and MOSE [6] for video segmentation tasks. |
| Dataset Splits | Yes | For video segmentation, we randomly select 100 videos and sample 15 consecutive frames from each for evaluation. For image segmentation, we randomly choose 50 videos and uniformly sample a total of 500 frames. |
| Hardware Specification | Yes | We conduct experiments on a machine with two NVIDIA A100-SXM4 GPUs, two Intel(R) Xeon(R) Gold 6132 CPUs and 314GB RAM. |
| Software Dependencies | No | The paper does not explicitly state specific software dependencies with version numbers (e.g., Python version, PyTorch/TensorFlow version, CUDA version). |
| Experiment Setup | Yes | We set the perturbation bound ϵ of the universal adversarial attack UAP-SAM2 to 10/255, and that of the sample-wise variant attack UAP-SAM2 to 8/255, using a batch size of 1 and training for 10 epochs. We use a fixed random seed of 30 for all experiments to ensure reproducibility. The attack achieves optimal performance under both settings when m = 256, which we adopt as the default configuration. The results shown in Fig. 7 (e) indicate that the attack performance stabilizes after the number of iterations reaches 10. Therefore, we set it as the default configuration for our experiments. Considering both computational efficiency and attack effectiveness, we set 30 as our default testing. Therefore, for efficiency considerations, we set 15 frames as the default configuration for our experiments. |