How to Build Trustworthy AI with Open Source
summary
Practical recipe for trustworthy AI using open source tools: a convolutional neural network on MNIST is shown with high accuracy but easily flipped from class 3 to 5 by a tiny perturbation, exposing a real gap between performance and safety. Demos use AI Explainability 360 with a contrastive explanations method to produce pertinent positives and negatives, and the Adversarial Robustness Toolbox to run a white-box evasion attack. Defenses like input compression and adversarial training are mentioned but not validated end to end.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:00AI has a great potential
- 1:45Are AI riskfree
- 4:04Ethical Problems
- 7:09Explainable AI
- 8:29Security and privacy vulnerabilities
- 10:31Trustworthy AI
- 11:31Linux Validation AI
- 12:00How to Achieve Trustworthy AI
- 14:00Exercise
- 15:12Getting familiar with the ecosystem
- 16:09Project categories
- 17:00Task definition
- 17:53Task explanation
- 19:47Robustness adversary evaluation
- 21:29Notebook
- 22:48Neural Network
- 24:18Results
- 25:32adversarial example
related talks
07mKeynote: Trusted AI: Bringing Trust Back into AI through Open Source
24mHow to Build Trustworthy AI
28mTrusted AI – Building Reproducible, Unbiased and...
30mTrustworthy AI - opening up the black box
46mCan you trust your AI?
38mTrustworthy AI: The Key to Unlocking High-Stakes Applications
17mAI Safety in the Generative AI Era
39mTrust Cards For AI