[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-501579-105":59,"doc-detail-501579-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","boosting-adversarial-robustness-using-feature-level-stochastic-smoothing","Boosting Adversarial Robustness using Feature Level Stochastic Smoothing","","Advances in adversarial defenses have significantly improved Deep Neural Network robustness, yet state-of-the-art methods still fall short for safety-critical robotics and autonomous navigation. Practical deployment also benefits from confidence estimation and the ability to reject unreliable predictions. This work introduces Feature Level Stochastic Smoothing to inject stochasticity into network predictions, smooth decision boundaries, and reject low-confidence outputs. The method boosts robustness on accepted samples, improves over existing adversarial training without rejection, and can be combined with adversarial detection to gain complementary benefits.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/boosting-adversarial-robustness-using-feature-level-stochastic-smoothing/501579/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/boosting-adversarial-robustness-using-feature-level-stochastic-smoothing/501579.png","ImageObject",300,407,{"name":92,"@type":93},"dabgoudw","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-22","2026-10-03",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",11,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does the paper address in adversarial robustness?","Question",{"text":113,"@type":114},"It targets the gap between improved robustness from existing defenses and the stronger requirements of safety-critical applications like robotics and autonomous navigation, including the need for reliable confidence and rejection of low-confidence predictions.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does Feature Level Stochastic Smoothing improve robustness?",{"text":118,"@type":114},"It introduces stochasticity in network predictions, smooths decision boundaries, and uses this mechanism to reject low-confidence predictions, boosting robustness on accepted samples.",{"name":120,"@type":111,"acceptedAnswer":121},"How does the method relate to adversarial detection and CCAT?",{"text":122,"@type":114},"It can be combined with confidence-calibrated adversarial training (CCAT) and detection methods to combine the benefits of adversarial training and detection while mitigating limitations of detection-only approaches.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},501579,1791040859,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":52,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":144,"read_time":145},3985747860484,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","Boosting Adversarial Robustness using Feature Level Stochastic Smoothing  \nSravanti Addepalli*, Samyak Jain*, Gaurang Sriramanan*, R. Venkatesh Babu  \nVideo Analytics Lab, Department of Computational and Data Sciences Indian Institute of Science, Bangalore, India  \nAbstract  \nAdvances in adversarial defenses have led to a significant improvement in the robustness of Deep Neural Networks. However, the robust accuracy of present state-ofthe-art defenses is far from the requirements in critical applications such as robotics and autonomous navigation systems. Further, in practical use cases, network prediction alone might not suf􀀂ce, and assignment of a con􀀂dence value for the prediction can prove crucial. In this work, we propose a generic method for introducing stochasticity in the network predictions, and utilize this for smoothing decision boundaries and rejecting low con􀀂dence predictions, thereby boosting the robustness on accepted samples. The proposed Feature Level Stochastic Smoothing based classi􀀂cation also results in a boost in robustness without rejection over existing adversarial training methods. Finally, we combine the proposed method with adversarial detection methods, to achieve the bene􀀂ts of both approaches.  \n1. Introduction  \nDeep Neural Networks are susceptible to carefully crafted imperceptible noise known as adversarial attacks [13], which can 􀀃ip their predictions to completely unrelated classes with high con􀀂dence. The catastrophic impact of such attacks has led to signi􀀂cant interest towards building defenses against such attacks.  \nAdversarial training using Projected Gradient Descent (PGD), proposed by Madry et [al](al. in)[. in](al. in) 2018 [20] has been one of the most successful defenses so far. PGD adversarial training coupled with early stopping is still one of the leading defenses [23, 21] indicating that progress on the front of adversarial defenses has been meager since 2018. Schmidt et al. [26] show that adversarial training requires signi􀀂cantly more data when compared to standard training. On the CIFAR-10 dataset for example, Carmon et al.  \n[7] demonstrate that a 7% increase in robustness requires 500K samples in addition to the 50K training samples of  \n* Equal contribution.  \nCorrespondence to: Sravanti Addepalli ([sravantia@iisc.ac.in](sravantia@iisc.ac.in))  \nthe original dataset. However, it is not practical to assume the availability of 10× more data for training robust models.  \nAnother avenue of research has been towards detecting adversarial samples [34, 24, 16, 12] . Such methods can be used to detect whether a test sample is adversarial or not, allowing the system to abstain from prediction on adversarial samples. However, most of the detection methods rely on identifying speci􀀂c properties of adversarial images. Hence, they work well under a black-box setting, but fail in the presence of an adaptive adversary, whose goal is to craft an adversarial example which is similar to the distribution of natural images, speci􀀂cally with respect to the property that is used for detection [31] .  \nRecent work by Stutz et al. [30] demonstrates a method of detecting different types of adversaries using Con􀀂dence Calibrated Adversarial Training (CCAT) . The authors propose to train networks that assign very low con􀀂dence to adversarial samples. This method is shown to induce lowcon􀀂dence predictions to adversaries constrained in other ℓp-norm balls as well, and demonstrates remarkably high detection accuracies, while limiting the rejection rate on correctly classi􀀂ed clean samples to 1% . However, this method is overly sensitive to even small random perturbations, which could occur due to factors such as slight soiling on a sensor collecting data. Based on our evaluation on the CIFAR-10 dataset, CCAT rejects 78.60% of images corrupted with Bernoulli noise of magnitude 1/255, while a model with standard training has an accuracy of 92 .4% on the same (Ref: Table-3 of the Supplementary) . Thus,","cbCaitp10U7cMTXm","https://ap.wps.com/l/cbCaitp10U7cMTXm","pdf",311306,"English","# Introduction\n## Motivation and limitations of existing defenses\n## Adversarial detection and CCAT shortcomings\n## Unified goal for robustness with selective rejection\n## Key contributions","[{\"question\":\"What problem does the paper address in adversarial robustness?\",\"answer\":\"It targets the gap between improved robustness from existing defenses and the stronger requirements of safety-critical applications like robotics and autonomous navigation, including the need for reliable confidence and rejection of low-confidence predictions.\"},{\"question\":\"How does Feature Level Stochastic Smoothing improve robustness?\",\"answer\":\"It introduces stochasticity in network predictions, smooths decision boundaries, and uses this mechanism to reject low-confidence predictions, boosting robustness on accepted samples.\"},{\"question\":\"How does the method relate to adversarial detection and CCAT?\",\"answer\":\"It can be combined with confidence-calibrated adversarial training (CCAT) and detection methods to combine the benefits of adversarial training and detection while mitigating limitations of detection-only approaches.\"}]","Boosting Adversarial Robustness using Feature Level Stochastic Smoothing | PDF",1791020893,25]