Securing AI Pipelines: Real-World Attacks on Kubernetes-Based AI Infrastructure
summary
Deploys Stable Diffusion from Hugging Face onto AKS using Kubeflow pipelines built from Jupyter notebooks (kfp, diffusers, transformers, torch), stores models and outputs in MinIO object storage, and runs a three-step pipeline: prepare/upload, generate image (15 inference steps, seed, CFG guidance scale), and promote deployment profile. Shows Kubernetes NetworkPolicy alone does not block L7 traffic and fixes it with Istio AuthorizationPolicy on service accounts, then exposes a Streamlit app on :8501 with rate limiting (≈5/min), regex/Chroma content filtering and per-user isolation; end-to-end demo succeeds.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 2:29Stable Diffusion architecture
- 6:04Kubeflow and ML workflows
- 8:11Kubeflow pipelines on Kubernetes
- 11:21Securing Kubeflow authentication
- 18:32Network policies for ML pipelines
- 21:42Secure text-to-image app
- 24:10Demo wrap-up
related talks
17mDevOps for Private AI: A Scalable Platform Using Kubeflow
36mShift-Left for LLMs: Securing the AI Model Supply Chain
46mDeploying scalable Machine Learning models with Kubernetes
42mKubernetes for Machine Learning: Managing and Scaling AI Workloads
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
30mOwned by Statistics: Using Kubeflow to Defend vs Attacks on Your ML Models
54mEconomics and Best Practices of Running AI/ML Workloads on Kubernetes
25mWhat We Learned Securing AI Agents at Scale on Multi-Tenant Kubernetes Clusters