当前位置:首页 > 报告详情

监视监视者——人工智能系统的保障和安全.pdf

上传人: 可*** 编号:991879 2025-12-07 29页 2.06MB

1、Watching the Watchers Watching the Watchers Safeguards and Security for Artificial Safeguards and Security for Artificial Intelligence systemsIntelligence systemsEoin Wickens-Director of Threat Intelligence“Nothing brings more peace of mind“Nothing brings more peace of mindthan than blind trust”blin

2、d trust”-Cybersecurity professionals nowhereArtificial Intelligence will Artificial Intelligence will replace us replace us allall-Artificial Intelligence(?)AI wont replace us AI wont replace us-but it will:but it will:Augment our workflowsAugment our workflows Automate Automate manymanytaskstasks E

3、mpower us to perform actions at scaleEmpower us to perform actions at scale Reduce cognitive loadReduce cognitive loadWe We blindly trustblindly trustAI models to perform actions on AI models to perform actions on our behalf and ultimately,our behalf and ultimately,decide our fatedecide our fate.AI

4、is only getting AI is only getting more capablemore capable,and is able,and is able to do to do more,with less,at scalemore,with less,at scale.Trust,but verify.Trust,but verify.How?How?PrePre-DeploymentDeployment(and the ML supply chain)(and the ML supply chain)PostPost-deploymentdeployment(Inferenc

5、e time attacks)(Inference time attacks)The AI development pipelineThe AI development pipelineTrainingTrainingProductionProductionThe AI development pipelineThe AI development pipelinePrePre-deploymentdeployment(Supply chain attacks)(Supply chain attacks)poisoningpoisoningbackdooring/hijackingThe AI

6、development pipelineThe AI development pipelinePostPost-deploymentdeployment(Inference attacks)(Inference attacks)adversarial examples,prompt injectionPrePre-DeploymentDeployment(and the ML supply chain)(and the ML supply chain)FUNDAMENTALS OF THEAI SUPPLY CHAINDATAMODELSTOOLINGINFRASTRUCTURESupply

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
根据报告的内容,全文主要内容概括如下: - **AI发展及其影响**: - AI将增强我们的工作流程、自动化许多任务、放大我们的行动规模并减轻认知负担。 - AI模型的能力不断增强,能够以更少的人力完成更多工作。 - **AI安全挑战**: - 预部署阶段(ML供应链)存在数据中毒、后门和供应链攻击风险。 - 部署后(推理时间攻击)面临对抗性样本、提示注入等威胁。 - **安全措施**: - 扫描威胁、签名验证、加密签名、安全评估。 - 确保模型艺术品的完整性,防止攻击。 - **关键数据**: - Hugging Face上有1.5M个模型,其中约40%的PyTorch/pickle模型存在高度安全风险。 - **关键点**: - 扫描、签名、验证。 - 模型存在于其系统环境中,需相应地威胁建模。 - 随着风险增加,需要更强大的安全控制。
如何防范模型攻击?" 揭秘潜在风险!" 安全监控指南!"
客服
商务合作
小程序
服务号
折叠