{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/generalizable-data-free-objective-for","title":"Generalizable Data-free Objective for Crafting Universal Adversarial Perturbations","arxiv_id":"1801.08092","date":"2018-01-24","proceeding":null,"authors":["Konda Reddy Mopuri","Aditya Ganeshan","R. Venkatesh Babu"],"abstract":"Machine learning models are susceptible to adversarial perturbations: small\nchanges to input that can cause large changes in output. It is also\ndemonstrated that there exist input-agnostic perturbations, called universal\nadversarial perturbations, which can change the inference of target model on\nmost of the data samples. However, existing methods to craft universal\nperturbations are (i) task specific, (ii) require samples from the training\ndata distribution, and (iii) perform complex optimizations. Additionally,\nbecause of the data dependence, fooling ability of the crafted perturbations is\nproportional to the available training data. In this paper, we present a novel,\ngeneralizable and data-free approaches for crafting universal adversarial\nperturbations. Independent of the underlying task, our objective achieves\nfooling via corrupting the extracted features at multiple layers. Therefore,\nthe proposed objective is generalizable to craft image-agnostic perturbations\nacross multiple vision tasks such as object recognition, semantic segmentation,\nand depth estimation. In the practical setting of black-box attack scenario\n(when the attacker does not have access to the target model and it's training\ndata), we show that our objective outperforms the data dependent objectives to\nfool the learned models. Further, via exploiting simple priors related to the\ndata distribution, our objective remarkably boosts the fooling ability of the\ncrafted perturbations. Significant fooling rates achieved by our objective\nemphasize that the current deep learning models are now at an increased risk,\nsince our objective generalizes across multiple tasks without the requirement\nof training data for crafting the perturbations. To encourage reproducible\nresearch, we have released the codes for our proposed algorithm.","url_abs":"http://arxiv.org/abs/1801.08092v3","url_pdf":"http://arxiv.org/pdf/1801.08092v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"generalizable-data-free-objective-for","repo_url":"https://github.com/val-iisc/gd-uap","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"unanswered"}},{"paper_slug":"generalizable-data-free-objective-for","repo_url":"https://github.com/HanxunH/XTransferBench","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"generalizable-data-free-objective-for","repo_url":"https://github.com/psandovalsegura/pytorch-gd-uap","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":"adversarial-attack","task_name":"Adversarial Attack"},{"task_slug":"depth-estimation","task_name":"Depth Estimation"},{"task_slug":"object-recognition","task_name":"Object Recognition"},{"task_slug":"semantic-segmentation","task_name":"Semantic Segmentation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1801.08092","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}