🤖 machine learning

The Calibrated Deepfake Trust Score (CDTS): Competence-Coupled Trust Degradation Across Deepfake Detectors

Dit artikel introduceert het Calibrated Deepfake Trust Score (CDTS) framework, dat aantoont dat de kalibratie van deepfake-detectoren fundamenteel gekoppeld is aan discriminatieve competentie, waardoor competentie wordt vastgesteld als een cruciale, labelvrije metriek voor het waarborgen van betrouwbaarheid, even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even even

Oorspronkelijke auteurs: Md Anas Biswas

Gepubliceerd 2026-06-30

📖 6 min leestijd🧠 Diepgaand

CC BY 4.0

Oorspronkelijke auteurs: Md Anas Biswas

Oorspronkelijk artikel gelicentieerd onder CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ Dit is een AI-gegenereerde uitleg van het onderstaande artikel. Het is niet geschreven of goedgekeurd door de auteurs. Raadpleeg het oorspronkelijke artikel voor technische nauwkeurigheid. Lees de volledige disclaimer

Stel je voor dat je een beveiligingsbeambte hebt bij een museum. Zijn taak is om nep schilderijen te ontdekken. In het verleden gaven we alleen om het feit of hij het meestal goed had (nauwkeurigheid). Maar in de echte wereld hebben we niet alleen een "ja/nee"-antwoord nodig; we moeten weten hoeveel we zijn "ja" of "nee" kunnen vertrouwen.

Als de bewaker zegt: "Ik weet voor 99% zeker dat dit een vervalsing is", maar hij is eigenlijk aan het gokken, dan is dat gevaarlijk. Als hij vol vertrouwen naar een echt schilderij wijst en het een vervalsing noemt, kan het museum het misschien weggooien. Dit artikel introduceert een nieuwe manier om dat vertrouwen te meten, de Calibrated Deepfake Trust Score (CDTS).

Hier is de kernontdekking van het artikel, eenvoudig uitgelegd:

Het Grote Geheim: "Competentie" is de Meesterschakelaar

De auteurs ontdekten dat het vermogen van een detector om een betrouwbare score te geven volledig afhangt van één ding: hoe goed hij daadwerkelijk is in het herkennen van de specifieke vervalsing waar hij naar kijkt. Ze noemen dit "competentie".

Denk aan een weervoorspeller:

Hoge Competentie: De voorspeller is een expert op het gebied van lokale stormen. Wanneer zij zeggen: "Er is een kans van 90% op regen", kun je die getallen vertrouwen. Als zij 90% zeggen, zal het ook echt 90% van de tijd regenen.
Lage Competentie: De voorspeller raadt over een weerpatroon dat hij nog nooit heeft gezien (zoals een nieuw type storm). Zelfs als zij zeggen: "Ik weet voor 90% zeker dat het gaat regenen", zijn ze eigenlijk gewoon aan het gokken. Hun vertrouwen is een leugen.

Het artikel bewijst dat naarmate de vaardigheid (competentie) van een detector daalt, het vermogen om een betrouwbare score te geven instort. Het is niet dat de wiskunde kapot is; het is simpelweg dat de detector niet weet waar hij naar kijkt, waardoor zijn "vertrouwen" betekenisloos wordt.

De Vijf Vertrouwenssignalen Bewegen Allemaal Samen

De onderzoekers keken naar vijf verschillende manieren om te meten of een detector betrouwbaar is:

Kalibratie: Komt de waarschijnlijkheid overeen met de werkelijkheid? (bijv. Betekent "80% kans" ook echt 8 van de 10 keer?)
Rechtvaardigheid (Fairness): Is de score even betrouwbaar voor verschillende groepen mensen (bijv. verschillende huidskleuren of leeftijden)?
Uitleg (Explanation): Wanneer de detector naar een deel van de afbeelding wijst en zegt: "Dit is het vervalste deel", kijkt hij dan ook daadwerkelijk naar dat deel?
Monitoring: Kunnen we merken wanneer de detector moeite heeft zonder dat we het antwoordmodel nodig hebben?
Routing: Kunnen we beslissen wanneer we het antwoord van de detector negeren?

De Verrassing: Het artikel laat zien dat dit geen vijf aparte problemen zijn. Ze worden allemaal gestuurd door die ene "Meesterschakelaar" (Competentie).

Als de detector bekwaam is, zijn alle vijf de signalen goed. De uitleg is logisch, de rechtvaardigheid is aanwezig en de scores zijn accuraat.
Als de detector onbekwaam is (te maken heeft met een nieuw type vervalsing), falen alle vijf de signalen tegelijkertijd. De uitleg wordt onzinnig, de scores worden onbetrouwbaar en de rechtvaardigheid stort in.

De "Tijd"-Misvatting

De auteurs dachten oorspronkelijk dat naarmate de tijd verstreek en er nieuwe vervalsingen werden uitgevonden, het vertrouwen langzaam zou wegvloeien. Ze kwamen erachter dat dit niet waar is.

Analogie: Het gaat niet om de kalender; het gaat om de afstand.
Een detector faalt niet simpelweg omdat een vervalsing "nieuw" is. Hij faalt omdat de vervalsing te ver verwijderd is van wat de detector getraind is om te zien. Als een detector getraind is op "kleimodellen" en je laat hem een "plastic model" zien, faalt hij. Als je hem een "houten model" laat zien (dat hij kent), werkt hij. De "tijd" die het kostte om het plastic model te maken, maakt niet uit; alleen het gebrek aan ervaring van de detector met plastic doet er toe.

De Oplossing: De "Label-Vrije" Radar

In de echte wereld heb je vaak niet het "antwoordmodel" (labels) om te controleren of een detector het bij het juiste eind heeft. Je hebt alleen de output van de detector.

Het artikel laat zien dat je de competentie van een detector kunt inschatten door simpelweg te kijken naar hoe de scores zich gedragen, zonder dat je de waarheid hoeft te weten.
Analogie: Stel je de motor van een auto voor. Je hoeft de motorkap niet te openen om te weten dat de motor moeite heeft; je kunt naar het motorgeluid luisteren (de verdeling van de scores). Als de motor hapert (hoge onzekerheid), weet je dat de auto in de problemen zit.
De auteurs ontdekten dat Predictive Entropy (een chique manier om te zeggen: "hoe verward de detector lijkt") een betrouwbare radar is. Als de detector verward lijkt, weet je dat zijn vertrouwensscore onbetrouwbaar is, zelfs als je niet weet waarom.

De Praktische Oplossing: "Competentie-bewuste Routing"

Omdat we nu kunnen detecteren wanneer een detector moeite heeft (lage competentie), kunnen we veranderen hoe we hem gebruiken.

Oude Manier: Als de detector met veel vertrouwen "Vervalsing" zegt, blokkeren we de inhoud. (Probleem: Een verwarde detector kan met veel zelfvertrouwen fout zitten).
Nieuwe Manier (CDTS): Voordat we de detector vertrouwen, controleren we zijn "competentie-radar".
- Als de detector bekwaam is voor dit specifieke type vervalsing, vertrouwen we zijn score.
- Als de detector moeite heeft (lage competentie), onthouden we ons van een oordeel. We zeggen: "Ik weet het niet, laat een mens hier even naar kijken," of we sturen het door naar een ander systeem.
Dit voorkomt dat het systeem met vol vertrouwen fouten maakt bij nieuwe, lastige vervalsingen.

Samenvatting

Het artikel betoogt dat we moeten stoppen met deepfake-detectoren te behanden als simpele "ja/nee"-machines. In plaats daarvan moeten we ze behanden als vertrouwsinstrumenten die constant gecontroleerd moeten worden op hun competentie.

Als de detector goed is in zijn werk: Zijn scores, uitleg en rechtvaardigheid zijn allemaal betrouwbaar.
Als de detector slecht is in zijn werk: Alles stort tegelijkertijd in.
De Oplossing: Gebruik een "competentie-monitor" (zoals een radar) om te detecteren wanneer de detector buiten zijn comfortzone treedt, en stop met het vertrouwen in zijn scores wanneer dat gebeurt. Dit maakt het hele systeem veiliger en betrouwbaarder.

Technische Samenvatting: De Calibrated Deepfake Trust Score (CDTS)

1. Probleemstelling

Moderne deepfake-detectoren worden steeds vaker ingezet, niet als binaire classificaties, maar als vertrouwenssignalen binnen operationele pipelines voor contentmoderatie, herkomstverificatie en identiteitsauthenticatie. In deze contexten wordt de output-waarschijnlijkheid van een detector geïnterpreteerd als een mate van vertrouwen; een zelfverzekend verkeerd gelabelde asset (hoge waarschijnlijkheid, onjuiste label) verspreidt zich geruisloos door een systeem zonder een review te triggeren. Bijgevolg is kalibratie—de betrouwbaarheid van de waarschijnlijkheidschatting—net zo cruciaal als discriminerende nauwkeurigheid.

De bestaande literatuur optimaliseert echter primair voor nauwkeurigheid (AUC, detectieratio's) en behandelt kalibratie als een statisch verbeteringsprobleem op vaste benchmarks. Deze aanpak faalt bij het adresseren van hoe vertrouwenssignalen degraderen wanneer detectoren geconfronteerd worden met nieuwe generatoren, compressie-artefacten of demografische subgroepen. Het artikel stelt dat huidige instrumenten voor vertrouwen een gebrek hebben aan "Graceful Trust Degradation" (GTD): het vermogen om gekalibreerd te blijven, getrouwe verklaringen te geven en stille fouten te voorkomen naarmate de onderliggende capaciteit van de detector afneemt.

2. Methodologie

De auteurs introduceren de Calibrated Deepfake Trust Score (CDTS), een wrapper-mechanisme dat de ruwe logit van elke detector transformeert naar een gekalibreerde vertrouwensscore ( $T$ ), aangevuld met een label-vrije competentiemonitor.

Experimentele Opzet

Detectoren: Drie architecturen werden geëvalueerd: Xception en EfficientNet-B4 (Convolutional Neural Networks) en een CLIP Vision Transformer. Alle waren getraind op face-swap vervalsingen.
Datasets: FaceForensics++ (c23), Celeb-DF v2, de DF40 generator-suite (variërend van GAN tot diffusie-tijdperken), en een externe FaceForensics DeepFakeDetection (DFD) set voor out-of-distribution validatie.
Competentie-manipulatie: Om competentie te isoleren van dataset-identiteit, gebruikten de auteurs een op reenactment getrainde (FR) Xception checkpoint op face-swap data. Dit induceerde een lage discriminerende competentie (AUC < 0,67) op anders detecteerbare vervalsingen, waardoor een gecontroleerd "lage-competentie" regime ontstond.
Kalibratieprotocol: De studie maakt gebruik van een oracle protocol waarbij een calibrator wordt aangepast op een uit de dataset gehouden, identiteits-disjuncte helft van de gelabelde data van elke doelgenerator. Dit isoleert de vraag of een score kalibreerbaar is gegeven de competentie van de detector, los van de overdraagbaarheid van een enkele calibrator.
Metrieken:
- Competentie: Area Under the ROC Curve (AUC).
- Kalibratie: Equal-mass Expected Calibration Error (ECE) om binning-bias te vermijden.
- Verklarende Getrouwheid (Explanation Faithfulness): Een verzadigingsvrije, per-regio rangorde metriek (Spearman correlatie tussen Grad-CAM belangrijkheid en de werkelijke waarschijnlijkheidsdaling bij patch-verwijdering) om confounds veroorzaakt door score-verzadiging in competente detectoren te vermijden.
- Label-vrije Signalen: Predictieve entropie, score-dispersie en referentie-divergentie (KS, Wasserstein) berekend zonder grondwaarheid-labels.

3. Belangrijkste Bijdragen en Bevindingen

C1: De Competentie–Kalibratie Koppeling

De centrale bevinding is een sterke negatieve correlatie tussen de discriminerende competentie van een detector en zijn kalibratiefout. Naarmate de competentie (AUC) daalt, stijgt de kalibratiefout (ECE).

Bewijs: Deze koppeling houdt stand over 32 configuraties (gepoolde Pearson $r = -0,81$ ), verspreid over twee CNN's en een transformer.
Causaliteit: De relatie is niet louter temporeel of dataset-specifiek. Wanneer competentie direct wordt gemanipuleerd (met de FR checkpoint op FS data), verstevigt de koppeling ( $r = -0,94$ ).
Externe Validiteit: De koppeling repliceert op een vierde, uit de dataset gehouden dataset (DFD) waarop de detectoren nooit getraind hebben, waarbij de per-actor correlaties significant blijven ( $r = -0,89$ voor Xception).
Inzetbare Realiteit: Een enkele calibrator die is bevroren op in-domain data faalt precies op lage-competentie generatoren. De fout van deze ingezette calibrator volgt de competentie nog nauwer ( $r = -0,98$ ) dan het oracle protocol, wat bevestigt dat vertrouwensscores onbetrouwbaar zijn precies daar waar nieuwe generatoren verschijnen.

C2: Kalibratie-gelijkheid vs. Nauwkeurigheid-gelijkheid

Het artikel demonstreert dat kalibratie-gelijkheid (calibration equity) onderscheidend is van nauwkeurigheid-gelijkheid (accuracy equity).

Bevinding: Subgroepen kunnen even nauwkeurig zijn (vergelijkbare AUC) maar ongelijk gekalibreerd zijn. Bijvoorbeeld, huidskleur-subgroepen vertoonden een AUC-verschil van 0,0006 maar een paargewijs ECE-verschil van 0,0151.
Implicatie: Standaard fairness-audits die zich uitsluitend richten op foutpercentages missen materiële discrepanties in de betrouwbaarheid van vertrouwensscores over demografische groepen heen.

C3: Verklarende Getrouwheid is Competentie-gekoppeld

Gebruikmakend van een verzadigingsvrije metriek, tonen de auteurs aan dat de getrouwheid van verklaringen stijgt met competentie ( $r = +0,94$ ).

Mechanisme: Competente detectoren produceren saliency maps die de beslissingen werkelijk drijven. Incompetente detectoren produceren saliency maps die ongecorreleerd zijn met het bewijs dat wordt gebruikt, waardoor verklaringen ongetrouw zijn.

C4: Label-vrije Competentiemonitoring

Het artikel identificeert predictieve entropie als een robuust, label-vrij signaal voor het schatten van competentie.

Draagbaarheid: In tegen tegenstelling tot referentie-divergentie signalen (zoals de KS-afstand), die van teken veranderen tussen architecturen (positief op Xception, negatief op EfficientNet), behoudt predictieve entropie een consistente negatieve correlatie met competentie over alle drie de architecturen ( $r = -0,84, -0,38, -0,95$ ).
Nut: Dit maakt de detectie van kalibratierisico's op ongeziene generatoren mogelijk zonder dat er grondwaarheid-labels nodig zijn.

C5: Unificatie van Vertrouwenssignalen

Op convolutionele detectoren convergeren intrinsieke vertrouwenssignalen (kalibratie en verklarende getrouwheid) naar een enkele, competentie-gealigneerde factor (die ~90% van de variantie verklaart). Hoewel de volledige multi-signaal structuur (inclusclusief monitoring-signalen) architectuur-specifiek is, is de kernbevinding—dat competentie de gedeelde drijfveer is van vertrouwensdegradatie—generalisabel over CNN's en transformers.

C6: Competentie-bewuste Routing

Routing van bron-batches op basis van een label-vrije competentie-inschatting (score-dispersie) presteert beter dan confidence-gebaseerde routing.

Resultaat: Competentie-bewuste routing verlaagt de algehele Area Under the Risk-Coverage (AURC) curve, specifiek door de prestaties te verbeteren in de laag-tot-middel bereik regio waar verificatiesystemen moeten beslissen of ze moeten afzien van een oordeel.

4. Betekenis en Claims

Het artikel betoogt dat de betrouwbaarheid van een detector niet een verzameling van vijf onafhankelijke garanties is (kalibratie, gelijkheid, getrouwheid, etc.), maar georganiseerd is rond één meetbare factor: competentie.

Herformulering van het probleem: De auteurs verwerpen de "temporele conjectuur" (dat vertrouwen simpelweg in de loop van de tijd degradeert). In plaats daarvan stellen zij dat tijd slechts een proxy is voor de afstand tot de competentie-grens van de detector. De signalen treden gelijktijdig op omdat ze een gedeelde competentie-driver delen.
Actiegerichte inzichten: Vertrouwensscoring moet competentie-bewust zijn. Systemen moeten competentie schatten (via label-vrije entropie) en het vertrouwen conditioneren op deze competentie.
Omvang en Bescheidenheid: De auteurs beperken hun claims met betrekking tot de "unificerende factor" expliciet. Hoewel de competentie-kalibratie koppeling en de label-vrije monitoring generaliseren over alle drie de architecturen, is de volledige unificatie van alle vertrouwenssignalen in één enkele factor specifiek voor convolutionele backbones. Het artikel beweert niet dat competentie een universele enkele factor is voor alle mogelijke architecturen of faithfulness-methoden, maar dat het de gedeelde drijfveer is die aangetoond is over de geteste CNN's en de transformer.

De CDTS-wrapper wordt voorgesteld als het mechanisme om dit inzicht te operationaliseren, waardoor systemen in staat zijn om te "weten, tonen, waarschuwen, niet-discrimineren en routeren" op basis van de werkelijke capaciteit van de detector in plaats van de ruwe zekerheid (confidence).