人工智能世界中的數據安全.pptx

1、Abhishek Das-Founder,VP of EngineeringDhruv Jain-Founder,Chief Product OfficerData Security in the AI WorldIHELLO DAVE.AM HERE NOW.Only 1/3rd of AI projects have reached productionModel Output Accuracy:39%*Retool,2023 State of AI AdoptionTop pain points for AI appsData Security:33%Hallucinations:28%

2、LLMs are the new era for Natural Language ProcessingDeep Neural NetworksTransformer Architecture with Self-attentionToken Embeddings&Context Similarity OK!This is not a technical talk on LLMsFoundational(LLM)ModelsPre-trained Foundational Models1234Closed-source(e.g.GPT,Claude,PaLM)Open-source(e.g.L

3、lama)CustomCompetitive advantage Privacy/confidentialityModel behavior guardrailsSingle vs Multi-model Multi-modal Compound AI SystemsRecommendation systemsChatbotsKnowledge-base Q&ALeveraging Foundational Models for InferenceUsing Proprietary DataArchitectural PatternsUse-CasesPrompt EngineeringRAG

4、Fine-tuning23No proprietary dataNo updates/changes to model weightsUnlocking the value of your proprietary dataAdd training on domain-specific datasetUpdates/changes to model weightsAdd&continuously update domain-specific knowledge-baseNo updates/changes to model weights*Retool,2023 State of AI Adop

5、tion1Almost 75%of enterprises are looking to use RAG or Fine-Tuning architectures*75%RAG and Fine-tuning Architectures are the Growing TrendQuery ExpansionQuery RewritingRe-rankingAugmentationInference LayerUser/Application InterfacesRoutingQUERYPrompt Eng+AugmentationPost-processingLLMModel LayerRE

6、SPONSEData LayerQuery ExpansionQuery RewritingRe-rankingAugmentationInference LayerUser/Application InterfacesRoutingQUERYPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRESPONSEQuery ExpansionQuery RewritingRe-rankingAugmentationInference LayerUser/Application InterfacesRoutingQUERYPr

7、ompt Eng+AugmentationPost-processingLLMData LayerModel LayerRESPONSERetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataQuery ExpansionQuery RewritingRe-rankingAugmentationInference LayerUser/Application Interface

8、sRoutingQUERYPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRESPONSERetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataYour proprietary data powers the AIDataDataDataQuery ExpansionQuery RewritingR

9、e-rankingAugmentationRoutingPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataDataDataDataAI risks start with Data Risks 111Privacy Risks-OWASP10 LLM

10、06-DASF Risk 1.21Query ExpansionQuery RewritingRe-rankingAugmentationRoutingPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataDataDataDataAI risks st

11、art with Data Risks 112Training data poisoning-OWASP10 LLM03-DASF Risks 1.4,1.7,2.1,3.121Query ExpansionQuery RewritingRe-rankingAugmentationRoutingPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Trainin

12、g DatasetCache:Previous QueriesLLM Generated DataDataDataDataAI risks start with Data Risks 113Prompt Manipulation-OWASP10 LLM01-DASF Risks 9.12331Query ExpansionQuery RewritingRe-rankingAugmentationRoutingPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRetrieverVector IndexDocument ch

13、unksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataDataDataDataAI risks start with Data Risks 114Unauthorized Access-OWASP10 LLM06-DASF Risks 2.1,3.1,9.123341Query ExpansionQuery RewritingRe-rankingAugmentationRoutingPrompt Eng+AugmentationPost

14、-processingLLMData LayerModel LayerRetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataDataDataDataAI risks start with Data Risks 115Sensitive Data Exfiltration-OWASP10 LLM06-DASF Risks 3.1,9.123345551Query Expans

15、ionQuery RewritingRe-rankingAugmentationRoutingPrompt Eng+AugmentationPost-processingLLMData LayerModel LayerRetrieverVector IndexDocument chunksEmbedding ModelFeature EngineeringHigh-quality Training DatasetCache:Previous QueriesLLM Generated DataDataDataDataAI risks start with Data Risks 116Data Supply-chain Poisoning-OWASP10 LLM03-DASF Risks 1.1,1.7,3.123345556661Acante is addressing these Data Layer risks of AIThank Youabhishekacante.aidhruvacante.aiWed love to see you in the Exhibit Hallwww.acante.ai