Unwanted Learning: What It Is, Why It Matters, and What To Do About It
summary
Empirical results show that simple classifiers trained on one label routinely learns identities from gender models and extract many unlabeled attributes (race, glasses, emotion) usable via transfer learning or white‑box probing. Representation‑censoring and membership‑inference experiments use standard NN toolkits and show that suppressing a protected attribute typically censoring destroys accuracy—the target task collapses toward random. This enables stealth backdoors, federated‑learning leaks, and raises GDPR purpose/consent problems because models retain repurposable, privacy‑violating features even after data deletion.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 0:50Title
- 1:55Machine Learning
- 5:22Leakage
- 8:08Inside the model
- 12:08Transfer learning
- 18:28Inferring race
- 22:45GDPR
- 25:27Repurposing
- 26:05Purpose
- 27:04Bias and Privacy
- 28:54Race and Gender
- 31:01No Transfer Learning
- 34:11No Engineered Features
- 34:37Unintended Effects
- 39:51Taking guesses
- 41:31What happens now
- 45:56This is it
- 46:44How do people train
- 48:51Whats the procedure
- 51:11Overfitting
- 53:13OverParameterize
- 54:36Linear Separator
- 56:36Implications
- 57:59Regulation
related talks
15mLearning and Unlearning Your Data in Federated Settings
28mPrivacy preserving machine learning
18mRectifying Privacy and Efficacy Measurements in Machine Unlearning...
28mSecurity of Machine Learning
13mHow Cloud AI Leaks What It Learns
22mWith Great Training Comes Great Vulnerability: Practical Attacks against Transfer Learning
29mBalancing Thin Line Between Data Intelligence & Privacy
24mData Governance in the AI age