[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83969-en":3,"doc-seo-83969-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},83969,7971461740886,"Theodore","https://ap-avatar.wpscdn.com/davatar_3d24733baf745e90a7e4bdd5f77d97b2",8,"Research & Report","Full-range Binary Classifier Calibration for Stable Model Updates in Production","Detection models deployed in adversarial environments face rapidly drifting malicious distributions while benign data remains comparatively stable, forcing frequent retraining and redeployment. Retraining can shift prediction scores and disrupt downstream users that rely on stable operational metrics. The work introduces a calibration method that targets a consistent false-positive rate (FPR) contract across the entire output range, not class probability, while shipping a compact pipeline. Experiments show relative FPR error within 2.3% from 10% to 0.1% and 7.2% at 0.01% FPR, with artifacts under 200 KB across calibration sets from 1K to 10M benign samples.","arXiv :2607 .0548 1v 1 [ cs .CR] 6 Jul 2026  \nFull-range Binary Classifier Calibration for Stable Model Updates in Production  \nKonstantin Berlin  \nCisco AI Defense  \nAbstract  \nDetection models running in adversarial environments face a malicious distribution that drifts rapidly while the benign distribution stays comparatively stable, so teams retrain andredeploy constantly to stay ahead of new threats. Retraining tends to change the output prediction scores, which breaks downstream users of the model. For these security-oriented models we need consistent false-positive rate (FPR) across all output values, whereas standard probability-calibration methods target class probability rather than an FPR contract. We introduce a method built on top of existing calibration primitives that targets the whole FPR curve, giving scores a consistent FPR meaning across deployments. On one held-out split, the observed relative FPR error was at most 2.3% from 10% down to 0.1% FPR and 7.2% at 0 .01% FPR. The shipped artifact remains under 200 KB in measurements across calibration sets from 1K to 10M benign samples.  \n1 Introduction  \nDetection models running in adversarial environments face a malicious distribution that drifts rapidly while the benign distribution stays comparatively stable. Teams therefore retrain and redeploy detection models continuously to stay ahead of evolving threats [1] . Retraining tends to change the output prediction scores, which breaks downstream users of the model. For these security-oriented models we need consistent false-positive performance across all output values. Standard probability-calibration methods such as Platt scaling and isotonic regression calibrate class probability rather than an FPR contract [2, 3] . We introduce a method that calibrates raw scores to the false-positive rate (FPR) directly on benign samples only, composing existing sklearn primitives (MinMaxScaler and IsotonicRegression) into a Pipeline that requires no custom inference code. The two-spline construction in Figure 1 keeps the FPR-to-rescaled-score lookup at fit time and ships only a raw-score-to-calibrated-score pipeline for production inference. Here, stability means that each model release is calibrated independently to the same fixed FPR-to-score contract. Raw-score thresholds may change after retraining, but a calibrated threshold retains the same target benign FPR.  \nThe FPR target depends only on the benign distribution. In the update setting we target, calibrating to FPRuses only benign traffic, which is the larger and more directly measurable sample pool, and avoids having to characterize adversarial behavior. In adversarial detection, positive examples are unknown unknowns because new attacks are hard to enumerate and label, yet their count remains far smaller than the benign count. FPR calibration is robust to positive-count uncertainty because it divides by the large, well-characterized benign count, whereas precision divides by the small, poorly characterized predicted-positive count, as quantified by the binomial planning rule in Section 2.  \nA single calibrated score can feed multiple downstream tiers (e.g., block at 0.1% FPR, alert at 1%, escalate at 10%), so calibration must hold across thresholds rather than only at one operating point.  \nThe method’s contributions are:  \n1. Whole-curve FPR mapping via a non-parametric monotone linear spline. Isotonic regression over the benign empirical CDF maps score thresholds to FPR at every threshold, with no assumed functional form.  \nFigure 1: Method overview. During fitting, the calibrator sorts benign scores, assigns plotting-position FPR labels, and fits a temporary FPR-to-rescaled-score spline. A fixed log-spaced FPR base grid augmented with that spline’s fitted edge knots then passes through the spline and the log-scale output contract to produce rescaled-score-to-calibrated-score storage knots. The second spline is the shipped sklearn artifact, so inference uses on","cbCaicVZVWPnKBLJ","https://ap.wps.com/l/cbCaicVZVWPnKBLJ","pdf",707839,3,1,10,"English","en",105,"# Introduction\n## FPR estimate bounds and biases\n## Sampling variance\n## Edge-of-sample bias","[{\"question\":\"Why is standard probability calibration insufficient for adversarial detection models?\",\"answer\":\"Standard methods calibrate class probability, while security systems require a consistent false-positive rate (FPR) contract across thresholds. Retraining changes prediction scores, which can break downstream users if FPR meaning is not preserved.\"},{\"question\":\"How does the proposed method ensure consistent FPR meaning across model releases?\",\"answer\":\"Each model release refits its calibration artifact on that release’s benign scores while keeping fixed FPR-to-score anchors. This preserves the same calibrated threshold contract even when raw-score thresholds shift after retraining.\"},{\"question\":\"What is the key idea behind the “whole-curve” calibration approach?\",\"answer\":\"The method maps raw scores to a calibrated scale that reproduces an FPR curve at every threshold. It uses isotonic regression over the benign empirical CDF and a two-spline construction, enabling consistent FPR interpretation across the full output range.\"}]",1784191729,25,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"full-range-binary-classifier-calibration-for-stable-model-updates-in-production","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":20},"https://docshare.wps.com/document/research-report/",{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/full-range-binary-classifier-calibration-for-stable-model-updates-in-production/83969/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-24","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why is standard probability calibration insufficient for adversarial detection models?","Question",{"text":75,"@type":76},"Standard methods calibrate class probability, while security systems require a consistent false-positive rate (FPR) contract across thresholds. Retraining changes prediction scores, which can break downstream users if FPR meaning is not preserved.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the proposed method ensure consistent FPR meaning across model releases?",{"text":80,"@type":76},"Each model release refits its calibration artifact on that release’s benign scores while keeping fixed FPR-to-score anchors. This preserves the same calibrated threshold contract even when raw-score thresholds shift after retraining.",{"name":82,"@type":73,"acceptedAnswer":83},"What is the key idea behind the “whole-curve” calibration approach?",{"text":84,"@type":76},"The method maps raw scores to a calibrated scale that reproduces an FPR curve at every threshold. It uses isotonic regression over the benign empirical CDF and a two-spline construction, enabling consistent FPR interpretation across the full output range.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":22,"slug":133},"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]