人工智能：复杂算法与有效的数据保护监—偏见评估督.docx

SUPPORTPOOL

OFEXPERTSPROGRAMME

AI-ComplexAlgorithmsandeffectiveDataProtectionSupervision

Biasevaluation

byDr.KrisSHRISHAK

AI-ComplexAlgorithmsandeffectiveDataProtectionSupervision-Biasevaluation

2

As

AspartoftheSPEprogramme,theEDPBmaycommissioncontractorstoprovidereportsandtoolsonspecifictopics.

TheviewsexpressedinthedeliverablesarethoseoftheirauthorsandtheydonotnecessarilyreflecttheofficialpositionoftheEDPB.TheEDPBdoesnotguaranteetheaccuracyoftheinformationincludedinthedeliverables.NeithertheEDPBnoranypersonactingontheEDPB’sbehalfmaybeheldresponsibleforanyusethatmaybemadeoftheinformationcontainedinthedeliverables.

Someexcerptsmayberedactedorremovedfromthedeliverablesastheirpublicationwouldunderminetheprotectionoflegitimateinterests,including,interalia,theprivacyandintegrityofanindividualregardingtheprotectionofpersonaldatainaccordancewithRegulation(EU)2018/1725and/orthecommercialinterestsofanaturalorlegalperson.

AI-ComplexAlgorithmsandeffectiveDataProtectionSupervision-Biasevaluation

3

TABLEOFCONTENTS

1Stateoftheartforbiasevaluation 5

1.1Sourcesofbias 5

1.1.1Biasfromdata 5

1.1.2Algorithmbias 6

1.1.3Evaluationbias 6

1.1.4Sourcesofbiasinfacialrecognitiontechnology 7

1.1.5SourcesofbiasingenerativeAI 7

1.2Methodstoaddressbias 8

1.2.1Pre-processing 9

1.2.2In-processing 11

1.2.3Post-processing 11

1.2.4MethodsforgenerativeAI 12

2Toolsforbiasevaluation 13

2.1IBMAIF360 13

2.2Fairlearn 13

2.3HolisticAI 14

2.4Aequitas 14

2.5What-IfTool 14

2.6Othertoolsconsidered 15

Conclusion 15

Bibliography 16

Documentsubmitt