Skupina STRADE prezentuje na konferenci IJCB 2026 (IEEE International Joint Conference on Biometrics), jedn\u00e9 z p\u0159edn\u00edch konferenc\u00ed v oblasti biometrie, dva p\u0159\u00edsp\u011bvky. Oba se zab\u00fdvaj\u00ed rozpozn\u00e1v\u00e1n\u00edm identity pomoc\u00ed modern\u00edch neuronov\u00fdch architektur, ka\u017ed\u00fd se v\u0161ak v\u011bnuje odli\u0161n\u00e9 modalit\u011b: prvn\u00ed rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed na v\u00fdpo\u010detn\u011b omezen\u00fdch za\u0159\u00edzen\u00edch, druh\u00fd ov\u011b\u0159ov\u00e1n\u00ed autorstv\u00ed textu.<\/p>\n\n\n\n
Rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed pro koncov\u00e1 za\u0159\u00edzen\u00ed<\/strong><\/p>\n\n\n\n Nasazen\u00ed v\u00fdkonn\u00fdch model\u016f pro rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed na za\u0159\u00edzen\u00edch s omezen\u00fdmi v\u00fdpo\u010detn\u00edmi zdroji, jako jsou mobiln\u00ed telefony nebo bezpe\u010dnostn\u00ed kamery, je omezeno kapacitou odleh\u010den\u00fdch s\u00edt\u00ed. B\u011b\u017en\u011b pou\u017e\u00edvan\u00fdm \u0159e\u0161en\u00edm je destilace znalost\u00ed, p\u0159i n\u00ed\u017e jsou znalosti p\u0159en\u00e1\u0161eny z rozs\u00e1hl\u00e9ho u\u010ditelsk\u00e9ho modelu do men\u0161\u00edho studentsk\u00e9ho. Standardn\u00ed destilace v\u0161ak nut\u00ed studentsk\u00fd model napodobovat i nespolehliv\u00e9 v\u00fdstupy u\u010ditele a nezohled\u0148uje rozd\u00edlnou kvalitu vstupn\u00edch sn\u00edmk\u016f.<\/p>\n\n\n\n Navr\u017een\u00e1 metoda kombinuje techniku Grouped Knowledge Distillation s p\u0159\u00edstupem MagFace, kter\u00fd zohled\u0148uje kvalitu vzorku. Prvn\u00ed slo\u017eka filtruje neinformativn\u00ed logity na \u00farovni v\u00fdstupu, druh\u00e1 zak\u00f3duje kvalitu sn\u00edmku p\u0159\u00edmo do geometrie p\u0159\u00edznakov\u00e9ho prostoru. Ob\u011b slo\u017eky tak p\u016fsob\u00ed na komplement\u00e1rn\u00edch \u00farovn\u00edch tr\u00e9novac\u00edho procesu. V\u00fdsledn\u00fd studentsk\u00fd model dosahuje v\u00edce ne\u017e \u010dty\u0159n\u00e1sobn\u00e9 propustnosti p\u0159i inferenci ve srovn\u00e1n\u00ed s u\u010ditelsk\u00fdm modelem.<\/p>\n\n\n\n Nejlep\u0161\u00ed konfigurace dos\u00e1hla nejvy\u0161\u0161\u00ed p\u0159esnosti na datasetu CFP-FP mezi porovn\u00e1van\u00fdmi metodami (96,24 %) a srovnateln\u00fdch v\u00fdsledk\u016f na datasetu AgeDB-30 (96,78 %). Studentsk\u00fd model nav\u00edc p\u0159eb\u00edr\u00e1 schopnost u\u010ditele odhadovat kvalitu sn\u00edmku a p\u0159i\u0159azuje vy\u0161\u0161\u00ed hodnoty magnitudy p\u0159\u00edznak\u016f kvalitn\u011bj\u0161\u00edm sn\u00edmk\u016fm tv\u00e1\u0159\u00ed.<\/p>\n\n\n\n Ov\u011b\u0159ov\u00e1n\u00ed autorstv\u00ed textu<\/strong><\/p>\n\n\n\n Ov\u011b\u0159ov\u00e1n\u00ed autorstv\u00ed, tedy rozhodnut\u00ed, zda dva texty poch\u00e1zej\u00ed od t\u00e9ho\u017e autora, p\u0159edstavuje \u00falohu behavior\u00e1ln\u00ed biometrie. Z\u00e1sadn\u00edm omezen\u00edm sou\u010dasn\u00fdch syst\u00e9m\u016f je jejich implicitn\u00ed z\u00e1vislost na s\u00e9mantick\u00e9 podobnosti: \u010dasto zam\u011b\u0148uj\u00ed tematickou shodu za shodu autorskou, a maj\u00ed tak pot\u00ed\u017ee spr\u00e1vn\u011b vyhodnotit tematicky odli\u0161n\u00e9 texty t\u00e9ho\u017e autora, p\u0159\u00edpadn\u011b chybn\u011b spojuj\u00ed texty r\u016fzn\u00fdch autor\u016f se spole\u010dn\u00fdm t\u00e9matem.<\/p>\n\n\n\n Navr\u017een\u00fd p\u0159\u00edstup se odkl\u00e1n\u00ed od tradi\u010dn\u00edch \u0159e\u0161en\u00ed zalo\u017een\u00fdch na enkod\u00e9rech a vyu\u017e\u00edv\u00e1 architekturu typu decoder-only (Qwen3). Autoregresivn\u00ed p\u0159edtr\u00e9nov\u00e1n\u00ed dekod\u00e9ru na predikci n\u00e1sleduj\u00edc\u00edho tokenu vede model k zachycen\u00ed stylistick\u00fdch charakteristik autora, nikoli pouze obsahu textu. Oba texty jsou spojeny do jednoho vstupu, co\u017e modelu umo\u017e\u0148uje vz\u00e1jemnou interakci mezi dokumenty prost\u0159ednictv\u00edm kauz\u00e1ln\u00edho attention mechanismu. Model je dotr\u00e9nov\u00e1n v\u00fdhradn\u011b metodou LoRA aplikovanou na projek\u010dn\u00ed matice attention vrstev, zat\u00edmco zbyl\u00e9 v\u00e1hy z\u016fst\u00e1vaj\u00ed zmrazen\u00e9. Sou\u010d\u00e1st\u00ed p\u0159\u00edstupu je rovn\u011b\u017e strategie t\u011b\u017eby tr\u00e9novac\u00edch p\u00e1r\u016f zam\u011b\u0159en\u00e1 na obt\u00ed\u017en\u00e9 p\u0159\u00edpady, kter\u00e1 model vede k rozli\u0161ov\u00e1n\u00ed na z\u00e1klad\u011b stylu, nikoli t\u00e9matu.<\/p>\n\n\n\n Nejlep\u0161\u00ed konfigurace (Qwen3-4B) dos\u00e1hla hodnoty ROC-AUC 0,9903 na vlastn\u00edm testovac\u00edm datasetu a 0,9462 na cross-domain benchmarku PAN21, a to bez tr\u00e9nov\u00e1n\u00ed na datech tohoto benchmarku. V porovn\u00e1n\u00ed se sout\u011b\u017en\u00edmi syst\u00e9my PAN21 se navr\u017een\u00e9 \u0159e\u0161en\u00ed um\u00edstilo nad n\u011bkolika p\u0159\u00edstupy, kter\u00e9 m\u011bly na rozd\u00edl od n\u011bj k dispozici tr\u00e9novac\u00ed data dan\u00e9ho benchmarku.<\/p>\n\n\n\n V\u00fdznam<\/strong><\/p>\n\n\n\n Oba p\u0159\u00edsp\u011bvky ukazuj\u00ed, \u017ee modern\u00ed neuronov\u00e9 architektury lze efektivn\u011b adaptovat na n\u00e1ro\u010dn\u00e9 biometrick\u00e9 \u00falohy. V prvn\u00edm p\u0159\u00edpad\u011b umo\u017e\u0148uj\u00ed spolehliv\u00e9 rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed p\u0159\u00edmo na koncov\u00fdch za\u0159\u00edzen\u00edch bez z\u00e1sadn\u00edho kompromisu mezi rychlost\u00ed a p\u0159esnost\u00ed, ve druh\u00e9m p\u0159edstavuj\u00ed n\u00e1stroj pro ov\u011b\u0159ov\u00e1n\u00ed autorstv\u00ed, jeho\u017e v\u00fdznam roste s rozvojem generativn\u00ed um\u011bl\u00e9 inteligence, a to jak pro forenzn\u00ed anal\u00fdzu, tak pro ochranu digit\u00e1ln\u00ed identity.<\/p>\n\n\n\n Tom\u00e1\u0161 Goldmann, Samuel \u0160im\u00fan Samuel \u0160im\u00fan, Tom\u00e1\u0161 Goldmann IJCB 2026 \u2014 IEEE International Joint Conference on Biometrics Skupina STRADE prezentuje na konferenci IJCB 2026 (IEEE International Joint Conference on Biometrics), jedn\u00e9 z p\u0159edn\u00edch konferenc\u00ed v oblasti biometrie, dva p\u0159\u00edsp\u011bvky. Oba se zab\u00fdvaj\u00ed rozpozn\u00e1v\u00e1n\u00edm identity pomoc\u00ed modern\u00edch neuronov\u00fdch architektur, ka\u017ed\u00fd se v\u0161ak v\u011bnuje odli\u0161n\u00e9 modalit\u011b: prvn\u00ed rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed na v\u00fdpo\u010detn\u011b omezen\u00fdch za\u0159\u00edzen\u00edch, druh\u00fd ov\u011b\u0159ov\u00e1n\u00ed autorstv\u00ed textu. Rozpozn\u00e1v\u00e1n\u00ed tv\u00e1\u0159\u00ed […]<\/p>\n","protected":false},"author":3,"featured_media":1345,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"ub_ctt_via":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-1344","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-strade"],"featured_image_src":"https:\/\/strade.fit.vutbr.cz\/wp-content\/uploads\/2026\/08\/hard_negative_mining_v3-e1788153983853.png","author_info":{"display_name":"igoldmann","author_link":"https:\/\/strade.fit.vutbr.cz\/index.php\/author\/igoldmann\/"},"_links":{"self":[{"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/posts\/1344","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/comments?post=1344"}],"version-history":[{"count":2,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/posts\/1344\/revisions"}],"predecessor-version":[{"id":1347,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/posts\/1344\/revisions\/1347"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/media\/1345"}],"wp:attachment":[{"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/media?parent=1344"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/categories?post=1344"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/strade.fit.vutbr.cz\/index.php\/wp-json\/wp\/v2\/tags?post=1344"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
\n\n\n\n
Magnitude-Aware Knowledge Distillation for Lightweight Face Recognition<\/em><\/p>\n\n\n\n
Semantics-agnostic Authorship Verification: Disentangling Style from Content using Specialized Decoder Architecture<\/em><\/p>\n\n\n\n
Podpo\u0159eno FIT VUT Brno, grant FIT-S-26-9011, a projektem e-INFRA CZ (ID:90254).<\/p>\n","protected":false},"excerpt":{"rendered":"