自治區公眾信任與安全維護的未來.pdf

《自治區公眾信任與安全維護的未來.pdf》由會員分享，可在線閱讀，更多相關《自治區公眾信任與安全維護的未來.pdf（15頁珍藏版）》請在三個皮匠報告上搜索。

1、August 2024AIADmitre.org 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.The ChallengeConsequential AI use can lead to national impact and transform society for the better but also poses systematic risksCatalyze consequential AI use 2024 T

2、HE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.The Need for the AI Assurance and Discovery Lab Mission needs necessitate rapid AI adoption to increase effectiveness and efficiency with limited resources Hype around tech breakthroughs complicate

3、 risk assessments and adoption decisions AI advancements can clash with human values and raise ethical concerns,which should be explored in controlled settings AI developers and governments are concerned about not fully understanding the capabilities of foundational models Regulators are under press

4、ure to act to ensure assured uses of AI without stifling innovation US Government is ordering assured uses of AI Executive Order 14110(October 30,2023)OMB Memo M-24-10(March 28,2024)2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.4Proactiv

5、ely discover and mitigate mission-critical risks in AI-enabled systems.Lab MissionLab VisionMission owners comprehend the risks of AI-enabled systems,make informed AI adoption decisions,and manage risks accordingly to maximize value from AI.Labs across the nation form a network of AI Assurance capab

6、ilities for public good.2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.5MITREs Definition of AI AssuranceAI assurance as a process for discovering,assessing,and managing risk throughout the lifecycle of an AI-enabled system so that it ope

7、rates effectively to the benefit of its stakeholders.AI ASSURANCE RISK GROUPSSecure:Cannot be tampered with,stolen,or easily circumventedEquitable:Does not promote harmful biasesInterpretable:Produces outputs that can be understood in a use contextRobust:Performs in varying conditionsPrivacy-enhance

8、d:Allows entities to control how their information is usedSafe:Does not endanger human life,health,property,or the environmentReliable:Performs consistently and is available when neededThere are many lenses for AI assurance and we adapt our approach to ensure it covers the risks that are important t

9、o the mission.AI Assurance Scope 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.6AI Assurance ServicesAI Assurance DiscoveryGiven an expected or imminent use case and the technology in an AI enabled system(AIES),exercise the risk landscap

10、e to understand the risks and value to the stakeholders of the AIES.Initialize an AI Assurance Plan coupled to the intended mission and set practical milestones for completing the AI Assurance Process.AI Assurance EvaluationEvaluate the risks and value of implementing the AIES to understand their li

11、kelihood and severity of impacts.Reaching quantitative results may entail hosting a sandbox at MITRE,validation testing against internal or synthesized datasets,and human-in-the-loop exploration.AI Assurance ManagementBased on risk evaluation results,provide risk mitigation strategies,monitoring req

12、uirements,and detailed suggestions for AI Governance.The plan will include prescriptive guidance for inevitabilities of the AIES like model drift and require the continued execution of the AI Assurance Process to maintain AI Assurance.AI Assurance Plan Development 2024 THE MITRE CORPORATION.ALL RIGH

13、TS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.AI Assurance Process7MITREs AI Assurance ProcessPosition Paper AI Assurance:A Repeatable Process for Assuring AI-enabled Systems 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED

14、 24-01019-7.8AI Assurance ProtocolExperts implement the AI Assurance Processfor your project to manage risks in consequential applicationsCommercial-off-the-Shelf AI Assurance ToolExplorationThe COTS Exploration Protocol and Environments allow MITRE to investigate the value of commercially available

15、 AI Assurance toolsRisk Discovery Protocol for AI AssuranceProvides risk awarenessfor consequential applications of AINavigate the AI assurance landscape and compare/contrast the desired application with similar use cases to prioritize risksThis Photo by Unknown Author is licensed under CC BY-SA-NCH

16、uman-in-the-loop Experimentation EnvironmentComplexity of human-AI interactions is driving new methods of measurement.The Lab provides two fully instrumented experiment rooms for simulating and observing human-AI interactions in mission contexts that can be synchronized with the AI Test HarnessHuman

17、-Centered AI Test HarnessAI-enabled systems need to work with humans so we need a capability for experimenting with human-AI interaction The AI Test Harness is a portable,web-based automated measurement platform for human-in-the-loop research and evaluationLab AI Assurance Capabilities(1/3)2024 THE

18、MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.9Lab AI Assurance Capabilities(2/3)AI Assurance Knowledge BaseProvides information to an AI assurance investigator on AI Assurance use cases,metrics,datasets,methodologies,and tools that are relevant

19、to their assurance goals.This Photo by Unknown Author is licensed under CC BYAI Red Teaming GuideBest practices on how to conduct AI red teaming,an investigative process that simulates adverse conditions on real-world AI enabled systems to identify vulnerabilities,mitigate potential exploits,and imp

20、rove the overall security posture and robustness of an AI-enabled system.Assurance Plan Templates and Development ProtocolsTools to facilitate the creation of an assurance plan and adoption of a development plan that will result in an assured AI-enabled system.Acquisition RFI Analysis ToolAn LLM-ena

21、bled tool that helps acquisition staff better understand and process RFIs and their responses.As such,the tool can be used by experts to identify,analyze,and augment RFI sections specific to AI assurance that should be driving AI-enabled system acquisitions.This Photo by Unknown Author is licensed u

22、nder CC BY-SA 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.10AI Governance ToolkitUnderstanding what is necessary to maintain an accountable AI system that proactively supports its mission,the Lab works with the AI Governance team to bu

23、ild comprehensive Assurance Plans that can be governedLab AI Assurance Capabilities(3/3)Adversarial Threat Landscape for AI Systems(ATLAS)A globally accessible,living knowledge base of adversary tactics and techniques based on real-world attack observations and realistic demonstrations from AI red t

24、eams and security groups,the Lab leverages ATLAS to assist in red teaming and risk discoveryLarge Language Model(LLM)Secure Integrated Research Environment(SIREN)Provides an environment to execute research and rapidly prototype assured LLM-based solutions aligned to mission use-cases,allowing the La

25、b to safely and securely work the AI Assurance Process with LLMs 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.11AI PlatformVirtualized GPU cluster for AI development,prototyping,and deploymentHPCsGPU enabled job cluster with adjacent st

26、orage for training dataFederal AI Sandbox248 H100 GPUs(1 exaFLOPS)that enables the training of Federal foundation modelsThis Photo by Unknown Author is licensed under CC BYThis Photo by Unknown Author is licensed under CC BY-NC-NDThis Photo by Unknown Author is licensed under CC BY-NC-NDCompute Reso

27、urces Available to the AIAD Lab 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.12 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.12AI-Enabled Augmented Reality MicroscopeTo inc

28、rease accuracy and reduce the time required for each cancer diagnosis,a commercial company integrated AI into a microscope,displaying results with an augmented reality interface.Our team investigated common microscopy-related pathology activities and developed software that alters imagery to evaluat

29、e the performance impact during real-world use.This included varying focus,lighting,chromatic aberration(colored distortion from lenses),and vignetting(a dark halo obscuring cells)which resulted in problems identifying potentially cancerous cells.The company was provided a report detailing risks and

30、 potential mitigations,with the suggestion that the technology be further refined before being used in a clinical setting.Completed Use Case1313ID VerificationAI can help expedite processing at airports,transportation hubs,and other environments where drivers licenses,passports,and other forms of ID

31、 are checked.Our assessment of ID verification systems identified several potential risks:Inconsistent capture of correct imagery of documents and facesUnequal face verification performance across varying demographic groupsLack of transparency about how verification decisions are madeMismatch betwee

32、n expectations and the reliability of automated authentication,leading to low utilization or over-relianceThe review also found that ID document review technologies are vulnerable to falsified data,which may be mitigated by implementing online database retrieval for high-security applications.Comple

33、ted Use Case 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.14Healthcare Mobile RobotDelegating routine,simpler tasks to autonomous systems in a healthcare setting lets providers focus on more complex work like diagnosing a patient or per

34、forming surgery.We procured a general-purpose robot platform and installed MITRE-developed software for autonomy and contact-less measurement of vital signs.The robot finds the patients room,verifies patient identity,gets into position to scan for vitals,records information,and returns to its starti

35、ng point.Completed Use Case 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-1018.We identified 58 risks and prioritized them.The two highest-priority hazards were:Patient misidentificationWe decreased misidentification risk by using multiple authe

36、ntication methods like those used in healthcare settings(e.g.,wristband checks,name confirmation,date of birth)Localization failure(getting lost)We added additional localization software to increase the robots ability to navigate correctly15AIAD Lab Ribbon Cutting25 March 2024Healthcare AI Assurance Lab at UMass from Blueprint 10 April 2024MITREs AIAD Lab as a Blueprint for other AI Assurance Labs 2024 THE MITRE CORPORATION.ALL RIGHTS RESERVED.APPROVED FOR PUBLIC RELEASE.DISTRIBUTION UNLIMITED 24-01019-7.