1、Safety,Security and Privacy in the Age of Generative AI,Carlos SotoSeptember 17,2024,Outline,AI and Digital TwinsThemes affecting Safety,Security&Privacy in Generative AIScalability challengesData security and privacyReliability and trustworthinessAI ecosystemsAdversarial risksKey research areas and
2、 additional challenges,2,AI and Digital Twins,Roles of AI in Digital TwinsForward models(e.g.,simulation surrogates,direct data-driven models)or components of themData analysis,control and decision supportOrchestrators(e.g.,LLM OS)For virtual modules,or entirephysical+virtual loopAI may also be used
3、 to build or inform DT components,3,Data analysis,Control,Decision support,Virtual,Physicalsensors,actuators,etc.,4,Molten Salt Reactors(MSRs)are advanced nuclear reactors with unique security,safeguards challengesDigital Twin of MSR core and diversion scenarios modeled via multi-physicsIsotopic com
4、position evolution(in-core and off-gas)Emission gamma spectra(in-core and off-gas)Delayed neutron precursorsAI module predicts risk of diversionPredict diversion strategies and rates,An AI-driven Digital Twin Example,Tracked chemical species,geometry and representation of molten salt fast reactor(IN
5、L&BNL,2024),AI for Science&Security,Safety,Security and Trustworthiness are necessary for effective AI use in sciences and securityAI outcomes and productsCorrectness,verifiability,repeatability,uncertainty quantification,explainabilityAI data and applicationsUser&data privacy,securityAction and aut
6、onomy validationAI infrastructureHPC,networking,streaming data,edge computing,5,Generative AI,the new normal,Frozen,general modelsLimited fine-tuningPrompt engineeringIn-Context LearningFew-shot learningChain-of-thought,Tree-of-thought,etc.External resourcesRetrieval-augmented generation(RAG)Tool us