This image captures a detailed scientific poster presented at the Computer Vision and Pattern Recognition (CVPR) conference in Seattle, WA, June 17-21, 2024. The poster details an experiment using SuperPixel Gradient Verification (SPGV) and includes a section titled "Logits" that describes the methodology involving pixel gradients and proxy models with illustrative diagrams. Table 1 titled “Baseline Comparisons” summarizes the experimental results across different datasets such as CUB200-2011, Indoor Scenes, and Diabetic Retinopathy, comparing various methods including ZSDB3KD, DFMS, and more against the proposed SPSG (SuperPixel Gradient) method. SPSG (Ours) shows promising results across multiple metrics, outperforming baseline methods in accuracy and effectiveness. The “Conclusion” section highlights the superiority of the SPSG method over existing ones across various datasets, noting its robustness in hard-label query scenarios and adversarial attack settings, indicating practical utility and effectiveness. Additionally, the affiliations noted at the top include logos from two institutions, suggesting collaboration between them. The poster concludes with contact information for Yunlong Zhao for further inquiries. The presented research is supported by the National Science Foundation of China and associated projects. A hand pointing towards the "Logits" section and individuals' feet are visible in the foreground, indicating the interactive nature of the presentation. Text transcribed from the image: CENTRA EB Experiments 浙 UNIVERSITY ZHEJIANG 1891 UNIVERSITY CVPR. SEATTLE, WA JUNE 17-21, 2024 Logits Method (probability) CUB200 (10k) CUB200 (10k) Table 1 Baseline Comparisons Agreement CUB200 (10k) Indoor(10k) Acc Indoor(10k) ZSDB3KD Queries Indoor(10k) 51.47 Agreement Acc 1.00 49.32 Queries 1202k 0.75 DFMS 59.41 55.21 58.37 1112k 53.19 b.50 1021k 61.11 EDFBA 60.12 0.25 55.61 1009k 53.12 459k 0.00 61.17 60.42 DS 349k Proxy model Pixel Gradient 200 54.97 53.98 1077k 60.27 59.98 1021k 100 100 DFME 52.31 50.17 1489k 58.71 58.17 1339k 200 0 SPSG(Ours) 60.71 55.47 132k 58.81 57.99 137k Method (probability) Caltech256 (10k) Caltech256 (10k) Caltehch256 (10k) Diabetic5(10k) Diabetic5(10k) Diabetic5(10k) Agreement Acc Queries Agreement Acc Queries knockoff 51.47 49.32 10k 31.12 29.32 10k Active Thief 55.21 53.19 10k 32.21 31.19 10k Black-Box Dissector 55.61 53.12 150k 34.81 33.27 150k InverseNet 56.19 55.32 150k 35.67 34.02 150k SPSG(Ours) 60.71 55.47 132k 36.12 35.25 122k Update Lgrad 1.00 Proxy model 1.00 0.75 0.75 b.50 0.50 0.25 0.25 0.00 0.00 200 200 100 0 100 100 100 200 0 200 0 Purified Supelpixle Simulated Superpixel Gradient Gradient aces two challenges: uires extensive query volume has a significant variance. on query images to estimate the s of the image. threshold strategy to reduce nct modules: SuperPixel Gradient rification (SGP). each sample at a low query cost, ersarial attack monitoring e proxy model. Within this module, the query set undergo a denoising We employ four datasets for our experimentation: CUB-200-2011, Indoor Scenes, and Diabetic Retinopathy. Baselines are categorized by their need for real data. Data-free baselines include DFME, DS, DFMS, EDFBA, and ZSDB3KD, while data-driven baselines, requiring real data, include Knockoff, Active Thief, Black-Box Dissector, and InverseNet. Felzenszwalb Quickshift Slic Grid Fig. 3 Gradients With Different Superplxel method Obtained superpixel gradients through queries under quick- shift, felzenszwalb, slic, and grid segmentation. Conclusion SPSG significantly outperforms existing MS algorithms across various datasets, demonstrating its effectiveness even in hard-label query scenarios. Its success in adversarial attacks showcases practical utility, and its ability to evade Prada highlights its stealthiness. Contact Yunlong Zhao | Email zhaoyl741@csu.edu.cn cience Foundation of China (Grant No. U2368201), special fund of National Key Laboratory of Ni&Co Associated Minerals Resources Development and Comprehensive Utilization(GZSYS-KY-2022-018, GZSYS-KY for Distinguished Young Scholars of Hunan Province (NO. 2023JJ10080).