沙查尔·梅纳谢_从MLOps到MLOops暴露机器学习平台的攻击面.pdf

《沙查尔·梅纳谢_从MLOps到MLOops暴露机器学习平台的攻击面.pdf》由会员分享，可在线阅读，更多相关《沙查尔·梅纳谢_从MLOps到MLOops暴露机器学习平台的攻击面.pdf（49页珍藏版）》请在三个皮匠报告上搜索。

1、#BHUSA BlackHatEventsFrom From MLOpsMLOps to to MLOopsMLOopsExposing the Attack Surface of Machine Learning PlatformsExposing the Attack Surface of Machine Learning PlatformsSpeaker:Shachar Menashe#BHUSA BlackHatEventswhoami Shachar Menashe Classically-Binary reverse engineer In practice-Full-time C

2、VSS assigner:)Leading JFrogs security research teams 0-day,CVE,malware research Presenting recent research from our 0-day team Ori Hollander,Natan Nehorai,Uriya Yavnieli#BHUSA BlackHatEventsOrg High Value Targets#BHUSA BlackHatEventsThis talk Breaking down MLOps platforms to distinct features How ca

3、n each feature be attacked?Chaining MLOps attacks for total domination l33t“ML Worm”demo How to avoid these attacks#BHUSA BlackHatEventsWhat can MLOps do for YOUThe ML software supply chainPretrained ModelML PipelineModel RegistryModel Serving#BHUSA BlackHatEventsWhat can MLOps do for YOUML Pipeline

4、Data InputData CleaningPre-processingModel TrainingDeployment#BHUSA BlackHatEventsWhat can MLOps do for YOUdsl.pipeline(name=XGBoost Trainer,)def xgb_train_pipeline(output=gs:/your-gcs-bucket,project=your-gcp-project,train_data=gs:/ml-pipeline-playground/sfpd/train.csv,eval_data=gs:/ml-pipeline-play

5、ground/sfpd/eval.csv,.):._analyze_op=dataproc_analyze_op().after(_create_cluster_op).set_display_name(Analyzer)_transform_op=dataproc_transform_op().after(_analyze_op).set_display_name(Transformer)_train_op=dataproc_train_op().after(_transform_op).set_display_name(Trainer).#BHUSA BlackHatEventsWhat

6、can MLOps do for YOUCV_model 1.2My_dev_model 0.1ChatGPT 4.5#BHUSA BlackHatEventsWhat can MLOps do for YOUModel Registry#BHUSA BlackHatEventsWhat can MLOps do for YOUEmbeddingServing#BHUSA BlackHatEventsWhat can MLOps do for YOUEmbeddingServing$kubectl apply-f-ENDapiVersion:machinelearning.seldon.io/