協作的自動化決策：通過設計構建安全、隱私和人權.pdf

《協作的自動化決策：通過設計構建安全、隱私和人權.pdf》由會員分享，可在線閱讀，更多相關《協作的自動化決策：通過設計構建安全、隱私和人權.pdf（54頁珍藏版）》請在三個皮匠報告上搜索。

1、#CiscoLive#CiscoLiveLisa Bobbitt,Principal Engineer Privacy,Regulatory AffairsllbobbittBRKCOL-2330Building in Security,Privacy and Human Rights by DesignCollaborations Automated Decisions 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveEnter your personal notes hereCisco W

2、ebex App 3Questions?Use Cisco Webex App to chat with the speaker after the sessionFind this session in the Cisco Live Mobile AppClick“Join the Discussion”Install the Webex App or go directly to the Webex spaceEnter messages/questions in the Webex spaceHowWebex spaces will be moderated by the speaker

3、 until June 9,2023.12343https:/ 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicBRKCOL-2330#CiscoLive 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicAgendaThe Need for Responsible AI Security,Privacy,and Human Rights RisksResponsible AI Principles&FrameworkResponsi

4、ble AI By DesignApplying Responsible AI to Collaborations Automated DecisionsThe Value of Responsible AIBRKCOL-23304 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveArtificial Intelligence5BRKCOL-2330APITraining DataDecisions,Insights,PredictionsService/Product Functional

5、Intent Service/Product Result 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicWhat is Generative AI?Generative Artificial Intelligence(GAI)Generative Artificial Intelligence(GAI)describes algorithms that can be used describes algorithms that can be used to create new content,includi

6、ng to create new content,including audio,code,images,text,video audio,code,images,text,video GAI is a Machine learning(ML)type of GAI is a Machine learning(ML)type of AIAI AI models“learn”from data patterns AI models“learn”from data patterns(training data,inputs)without human(training data,inputs)wi

7、thout human direction direction Two common ML models used in GAI:Two common ML models used in GAI:1.1.Diffusion ModelsDiffusion Models For image generation tools like For image generation tools like Stable Diffusion and MidjourneyStable Diffusion and Midjourney2.2.Large Language Models(LLMs)Large La

8、nguage Models(LLMs)For tools like ChatGPT and CopilotFor tools like ChatGPT and CopilotBRKCOL-23306 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveGenerative Artificial Intelligence(AI)Models Democratized7BRKCOL-2330Training DataAudio Images Decisions Insights,Predictions

9、UICHATCHAT GPT3/4ApplicationsAPITextCode01001001 11 00PROMPTData Becomes Code 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicWhat about Model and Model and Training DataTraining DataRisks?Poor data quality Poor data selectionWrong outputsInstabilityLack of reproducibilityImproper a

10、pplicationConfirmation bias Concept drift/off-label useInadequate consideration of assumptions and limitations8BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicWhat about BusinessBusinessRisks?Exposure of customer/partner confidential dataExposure of company confidential

11、dataIP infringement(code)Copyright infringement(images and text)Loss of patents rights or copyrightOpen-Source contamination9BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicWhat about UserUserRisks?Violation of privacy lawsRisks to Human RightsSecurity vulnerabilitiesAcc

12、uracy and safety issuesLack of transparency/understandingLack of accountability 10BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBenefits:Virtual Assistants can increase meeting accessibility and efficiency in meetingsWebex Assistant and TranslationBenefits:Vi

13、rtual backgrounds can increase user privacy and representation in meetingsVirtual BackgroundsBenefits:Noise Removal increases user privacy,representation,and comfort in meetingsBackground Noise ReductionWebexs Automated Decisions.11BRKCOL-2330Benefits:Facial Recognition can increase identification o

14、f the speaker,aiding collaboration and representation in meetingsFacial Recognition Webex has multiple functions that make automated decisions for an individual to enhance the collaborative experience if if the individual wants to use themBenefits:Actionable Insights,Code Automation,Automated Chat S

15、ummariesWebex Contact Center and ConnectFor a deeper look at AI in Collaboration,check out Keiths BRKCOL-1871 Cognitive Collaboration 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLivePotential Risks of Collaboration Automated Decisions13BRKCOL-2330BusinessBusinessExposure o

16、f customer/partner confidential dataExposure of company confidential dataTraining data copyright infringement(images,video,audio,and text)IP infringement(code)Webex UsageWebex UsageRisks to privacy Risks to human rights Security vulnerabilitiesLack of transparency/understandingEnd users combined ris

17、ks Sensitive data environments:Education,Healthcare,JusticeHybrid experiences work,family,and home exposureModel and Training DataModel and Training DataPoor data quality Poor data selectionWrong outputsInstabilityImproper applicationConfirmation bias Concept drift/off-label useResponsible AI/ML 202

18、3 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveCiscos Responsible AI Principles15BRKCOL-2330FairnessFairnessSecuritySecurityAccountabilityAccountabilityTransparencyTransparencyReliabilityReliabilityPrivacyPrivacy 2023 Cisco and/or its affiliates.All rights reserved.Cisco Pub

19、lic#CiscoLiveTransparencyCiscos Responsible AI Principles(1/6)Ciscos goal is to provide clarity and consistency in informing users about our application of AI in a manner that is accessible,transparent,and understandable.This includes:oWhen AI is employed in our technologiesoThe intent of the AI and

20、 its model classoThe data demographicsoSecurity,privacy and human rights controls applied to the modeloHow to get more information about our use of AI16BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveFairnessCiscos Responsible AI Principles(2/6)Cisco strives to

21、 identify and remediate harmful bias within our algorithms,training data,and applications that are directly involved in consequential decisions.Consequential decisions are those that could have a legal Consequential decisions are those that could have a legal or human rights impact on individuals or

22、 groups.or human rights impact on individuals or groups.We have developed mechanisms for our customers to We have developed mechanisms for our customers to provide feedback and raise any concerns for review and provide feedback and raise any concerns for review and action by our Incident Response Te

23、am.action by our Incident Response Team.17BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveAccountabilityCiscos Responsible AI Principles(3/6)The Cisco Responsible AI Framework requires teams to account for privacy,security,and human rights impacts from the very

24、 beginning of development through the end of the AI lifecycle.Cisco is committed to upholding and respecting the human rights Cisco is committed to upholding and respecting the human rights of all people,as articulated in our of all people,as articulated in our Global Human Rights PolicyGlobal Human

25、 Rights Policy.Accountability measures include requiring documentation of AI Accountability measures include requiring documentation of AI use cases,conducting impact assessments,and oversight use cases,conducting impact assessments,and oversight provided by a group of crossprovided by a group of cr

26、oss-functional leaders.functional leaders.18BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLivePrivacyCiscos Responsible AI Principles(4/6)Cisco has built privacy engineering practices into the Cisco Secure Development Lifecycle(CSDL)to design,build,and operate pr

27、ivacy-enhancing features,functionality,and processes into our offers.These apply to training data,prompts,and results.When processing personal information,Cisco is committed When processing personal information,Cisco is committed to following the principles set forth in our Global Personal to follow

28、ing the principles set forth in our Global Personal Data Protection and Privacy Policy,which aligns with Data Protection and Privacy Policy,which aligns with applicable international privacy laws and standards.applicable international privacy laws and standards.19BRKCOL-2330 2023 Cisco and/or its af

29、filiates.All rights reserved.Cisco Public#CiscoLiveSecurityCiscos Responsible AI Principles(5/6)Cisco builds AI technologies using leading security practices,drawing on our secure development lifecycle to maximize resilience and trustworthiness.To meet the unique characteristics of AI,Cisco has adde

30、d To meet the unique characteristics of AI,Cisco has added specific security controls for AI that improve attack resiliency,specific security controls for AI that improve attack resiliency,data protection,privacy,threat modeling,monitoring,and data protection,privacy,threat modeling,monitoring,and t

31、hirdthird-party compliance.party compliance.20BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveReliability Ciscos Responsible AI Principles(6/6)Cisco designs and tests AI systems and their components for reliability.As part of our responsible AI assessment,we re

32、view As part of our responsible AI assessment,we review AIAI-based solutions for embedding controls in their lifecycle based solutions for embedding controls in their lifecycle to maintain consistency of purpose and intent when to maintain consistency of purpose and intent when operating in varying

33、conditions and use cases.operating in varying conditions and use cases.Where we identify that an AI solution has potential impacts Where we identify that an AI solution has potential impacts on safety,we impose additional integrity controls.on safety,we impose additional integrity controls.21BRKCOL-

34、2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveCiscos Responsible AI 22BRKCOL-2330PrinciplesFairness,Privacy,Security,Reliability,Accountability,TransparencyFrameworkGovernance,Controls,Incident Management,Industry Leadership,External EngagementResultsResponsible AI

35、By Design 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveIndustry LeadershipIncident ManagementControlsGovernance&OversightThe Responsible AI Framework operationalizes our principles throughout the company.Ciscos Responsible AI Framework23BRKCOL-2330External Engagement 20

36、23 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveGovernance&OversightEstablishes a Responsible AI Committee of senior executivesAdvises on Responsible AI practices and oversees Responsible AI Framework adoptionReviews high-risk applications of AI proposed by business units an

37、d incident reports 24BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveControls Embeds legal,security,privacy,and human rights processes Embeds legal,security,privacy,and human rights processes as part of theas part of the existing Cisco Secure Development Lifecy

38、cle intoexisting Cisco Secure Development Lifecycle intoInternally designed AI modelsInternally designed AI models3 3rdrd-party modelsparty modelsSelection of training data Selection of training data Tracking of use Tracking of use Assesses applications embedding AI for adverse impacts to Individual

39、s and/or groups of peopleCustomersCiscoApplies to reduce risk of harm,including legal,unintended bias,privacy,model monitoring,and transparency25BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveIncident Management Leverages security,data breach,and privacy incid

40、ent response system to manage reported AI incidents involving bias and discriminationEscalates incidents to the Responsible AI Incident Response Team to addressTracks and reports AI incidents and remediation to governance board and other relevant stakeholders26BRKCOL-2330 2023 Cisco and/or its affil

41、iates.All rights reserved.Cisco Public#CiscoLiveIndustry LeadershipEmbeds Responsible AI as a focus area for incubation of new Embeds Responsible AI as a focus area for incubation of new technology across Ciscotechnology across CiscoEngages with industry innovation providers focused on delivering Re

42、sponsible AIParticipates proactively in industry forums to advance Responsible AI,including the Centre for Information Policy Leadership,Equal AI,and Business Roundtable on Human Rights and AI27BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveExternal Engagement

43、 Works with governments to understand global perspectives on AIs Works with governments to understand global perspectives on AIs benefits and risks benefits and risks Monitors,tracks,and influences AI-related legislation,emerging policy,and regulationsPartners with and sponsors cutting-edge research

44、 institutions,exploring the intersection of ethics and AI from technical,organizational,social,and design perspectives28BRKCOL-2330Applying ResponsibleAI/ML By Design 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveResponsible AI By Design 30BRKCOL-2330 2023 Cisco and/or i

45、ts affiliates.All rights reserved.Cisco Public#CiscoLiveRAI Workflow31BRKCOL-2330AI Functional AI Functional Concept/IntentConcept/IntentModel SelectionModel SelectionData SelectionData SelectionSecurity,Privacy,and Security,Privacy,and Human Rights By Human Rights By Design:RAI Impact and Design:RA

46、I Impact and Risk AssessmentRisk AssessmentTreatments for Identified Treatments for Identified Risks by Applying Risks by Applying ControlsControlsBuild&Verify ModelBuild&Verify ModelEmbedded/Updated in Embedded/Updated in Cisco Offer and/or Cisco Offer and/or Enterprise Processing Enterprise Proces

47、sing ActivityActivity 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveResponsible AI Assessment32BRKCOL-2330Intended&Unintended UseUse CasesInternal,3rdParty,Rights&PermissionsModel InfoData Origin,Content,Retention,Aggregation,LabelingTraining DataLegal,Privacy,Fairness,S

48、ecurity,Reliability,Transparency,AccountabilityIdentify Risks&Apply TreatmentsRisk-Based Assessment with Cloud Control Framework RAI Controls to Lower Risk 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveSample Questions from the Assessment33BRKCOL-23301.What is your inten

49、t for this function?2.Do you have legal and commercial use rights?3.What use cases are explicitly out of scope?4.Does this model generate output that results in a consequential decision affecting a user or a certain group of users?5.Has this model been tested for differing outcomes by demographic ca

50、tegory?6.Does this model include a mechanism or process that enables feedback from a user?2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveAI/ML Incident Response34BRKCOL-2330Receive ML/AI Incident ReportsReceive ML/AI Defect Trend ReportsPSIRT management processEngineering

51、 defect tracking processIncident attack-relatedDefectRespond to defect reporterWhatis thisreport?PSIRT update industry and customers with status Field Advisory Customer Inquiry Clearinghouse Trust PortalEngineering defect tracking processIncident notnot attack-relatedApplying Responsible AI/ML into

52、Collaboration 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveTransparency-Webex AI Addendum to Privacy Data Sheet on 36BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveFairness Applied to Collaboration37BRKCOL-2330Assessments of the m

53、odel,its development,and its production environment for consequential decisions and result affecting human rights and privacy Mens versus womens voice rangeHead coverings and hair stylesCulturally inappropriate results from generative promptsUser control of the user of the AI functionsCompany determ

54、ines if the capability is turned onEnd user turns it on for themselves 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveWebex AI Accountability38BRKCOL-2330Design AccountabilityDesign AccountabilityWebex AI Facial Recognition is only used in the Collaboration Products as th

55、e end-user has control of its useEach data set used for training of Webex AI/MLgo through a review with both the Privacy Office and Product LegalOperational AccountabilityOperational AccountabilityWebex AI responses to feedback when end-user experiences do not align to their expectations 2023 Cisco

56、and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveLegal and Privacy Review of Training DataWebex Example39BRKCOL-2330Review includesLegal use of the dataCommercial use of the dataReview of PII that is in the data setRecommendations of minimization of PII through De-Identification,Anony

57、mization,and DeletionReview of security and access of the training data 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveWebex Security 40BRKCOL-2330Webex is designed and developed via the Cisco Secure Development LifecycleReview includesOperational Security Platform Securi

58、tySecure Data StorageSecure Data In TransitAccess management of customers use including all AI functionsFor security details: 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBenefitsBenefits:Virtual backgrounds can increase user privacy and representation in meetingsRisksR

59、isks:Early models did not perform as well for all hair textures,hairstyles,skin tones or lighting conditionsRemediationRemediation:Added more hair textures,styles,skin tones,and lighting conditions to training dataVirtual BackgroundsBenefitsBenefits:Noise Removal increases user privacy,representatio

60、n,and comfort in meetingsRisksRisks:Early models did not perform as well for higher-pitched voices RemediationRemediation:Created pitch-balanced test sets,added more high-pitch voices to training data,and expanded the subjective test suiteBackground Noise ReductionResponsible AI/ML in Webex41BRKCOL-

61、2330Responsible AI Impact Assessments focus on the potential impacts of intelligent product components but may not consider the cumulative impacts of those components.2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveResponsible AI/ML in WebexResponsible AI Impact Assessment

62、s focus on the risks of the AI42BRKCOL-2330BenefitsBenefits:Virtual Assistants can increase meeting accessibility and efficiency in meetingsRisksRisks:Virtual Assistants may not perform as well for all languages,dialects,accents,or pitches for transcription into captions and translation.Poor transcr

63、iption contributes to product inaccessibility.RemediationRemediation:Include diverse,high-quality training data appropriate for Webexs use casesWebex AssistantBenefitsBenefits:Individuals can be identified in a meeting without maintaining their imageRisksRisks:Use of facial recognition for other pur

64、posesRemediationRemediation:Limited access to the function to only Webex Facial RecognitionThe Value of Applying Responsible AI/ML 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicDoes your organization need to do more to reassure customers about their data for AI?Source:Cisco 2023 D

65、ata Privacy Benchmark Study92%Yes6%No2%UnsureBRKCOL-233044 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveFigure 8.Support and Concerns About the Use of Personal Data in AISupport for AI UseConcerns About Current AI UseBelieve AI can be useful in improving our lives43%Wil

66、ling to share anonymized personal data to improve AI products54%Concerned about the business use of AI today 60%Use of AI by organizations has already eroded trust in them65%Source:Cisco 2022 Consumer Privacy SurveyWhat about using AI and personal data?BRKCOL-233045 2023 Cisco and/or its affiliates.

67、All rights reserved.Cisco Public#CiscoLiveWhat makes consumers more comfortable with AISource:Cisco 2022 Consumer Privacy Survey36%30%28%28%23%16%40%45%46%47%46%49%24%25%26%25%31%35%0%10%20%30%40%50%60%70%80%90%100%Provide opportunity to opt outInstitute AI ethics management programExplain how appli

68、cation makes decisionsInvolve a human in decision-making processAudit for biasAdopt AI ethics principlesMuch More ComfortableMore ComfortableNo DifferenceBRKCOL-233046 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveWhat organizations are doingSource:Cisco 2023 Data Privac

69、y Benchmark StudyWhat organizations have doneWhat organizations say would be most effective21%47%53%55%60%63%0%10%20%30%40%50%60%70%Ensuring a human is involved in the processExplaining how the AI application worksAdopting AI ethics principlesApplying an AI ethics management program to identify and

70、reduce unintended biasAuditing for biasGiving customers the opportunity to opt out of the AI use22%41%43%53%55%58%0%10%20%30%40%50%60%70%Explaining how the AI application worksEnsuring a human is involved in the processAdopting AI ethics principlesApplying an AI ethics management program to identify

71、 and reduce unintended biasAuditing for biasGiving customers the opportunity to opt-out of the AI useBRKCOL-233047 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveMaintain Customer TrustCompete Deliver on Industry StandardsComply with Emerging RegulationsResponsible AI Ben

72、efits48BRKCOL-2330 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveResponsible AI/ML Resources49BRKCOL-2330The Cisco Responsible AI/ML FrameworkThe Cisco Responsible AI/ML F Principles for Responsible AICisco Principles for Responsible AI 2022 Consumer Privacy SurveyCisco

73、2022 Consumer Privacy S Is Key:Introducing Cisco Responsible AITransparency Is Key:Introducing Cisco Responsible AI Meeting on Cisco Trust Portal Privacy Data Sheet Webex Meeting on Cisco Trust Portal Privacy Data Sheet-Addendum for AI Functions Addendum for AI Functions 2023 Cisco and/or its affili

74、ates.All rights reserved.Cisco Public#CiscoLiveFill out your session surveys!Attendees who fill out a minimum of four session surveys and the overall event survey will get Cisco Live-branded socks(while supplies last)!50BRKCOL-2330These points help you get on the leaderboard and increase your chance

75、s of winning daily and grand prizesAttendees will also earn 100 points in the Cisco Live Game for every survey completed.2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicContinue your educationVisit the Cisco Showcase for related demosBook your one-on-oneMeet the Engineer meetingAtte

76、nd the interactive education with DevNet,Capture the Flag,and Walk-in LabsVisit the On-Demand Library for more sessions at www.CiscoL you#CiscoLive 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLive53Gamify your Cisco Live experience!Get points Get points for attending this

77、session!for attending this session!Open the Cisco Events App.Click on Cisco Live Challenge in the side menu.Click on View Your Badges at the top.Click the+at the bottom of the screen and scan the QR code:How:123453 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicBRKCOL-2330#CiscoLive