Defending Against Adversarial Model Attacks using Kubeflow
summary
Demonstrates concrete defenses by integrating the Adversarial Robustness Toolbox (ART) into Kubeflow Pipelines and KFServing, with payload logging for production monitoring. Shows FGSM that drops test accuracy from 87% to 13%, a PGD example that causes a classifier to call a cat image a “basketball,” and a Square attack that flips a handwritten 3 into a 9 with a small rectangle; retraining on collected adversarial examples restores robustness. Project has >150,000 downloads and live pipeline/explainer demos.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:03Presentation
- 3:07Vision for Trusted AI
- 15:41Adversarial Threats
- 22:39Art and Practice
- 25:01Kubeflow
- 28:19Kubeflow pipelines
- 36:10Square Attack
- 36:39Parameters
- 37:13Explainer Spec
- 37:48Explainer Flow
- 38:12Explainer Demo
- 40:54Adversarial Team
- 41:19Contact Information
- 41:35Conclusion
related talks
21mDefending 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...
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples
26mObjection! AI Security Mistakes on Trial With Kubeflow and Confidential Computing
40mMLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
38mUsing adversarial samples to robustify your Neural Network Models
21mAdversarial Attacks