Defending Against Adversarial Model Attacks Using Kubeflow
summary
Built pipelines that embed Adversarial Robustness Toolbox (ART) into Kubeflow Pipelines and a KFServing explainer to run attacks and defenses automatically. Components execute FGSM, PGD and Square attacks and produce metrics: baseline accuracy, accuracy on adversarial samples, confidence reduction on correctly classified adversarial samples, and average perturbation for misclassified samples. Demos show PGD dropping cat confidence 92%→4% (spatial smoothing recovers to 76–92%) and a Square attack turning an MNIST 3 into a 2; robustness flag returns false when vulnerable.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:45Trusted AI
- 3:40Open Governance
- 4:10Security
- 6:55Toolkit
- 8:55Other Projects
- 11:07Adversarial robustness toolbox
- 16:45Kubeflow Survey
- 17:53Demo
related talks
43mDefending Against Adversarial Model Attacks using Kubeflow
30mOwned by Statistics: Using Kubeflow to Defend vs Attacks on Your ML Models
29mOwned By Statistics: How Kubeflow & MLOps Can Help Secure...
40mMLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
38mUsing adversarial samples to robustify your Neural Network Models
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples
26mObjection! AI Security Mistakes on Trial With Kubeflow and Confidential Computing
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate