Can a Model Be Differentially Private and Fair?
Training models with differential privacy stops models from inadvertently leaking sensitive data, but there's an unexpected side-effect: reduced accuracy on underrepresented subgroups.
Can a Model Be Differentially Private and Fair? Explorables Can a Model Be Differentially Private and Fair? Training models with differential privacy stops models from inadvertently leaking sensitive data, but there's an unexpected side-effect: reduced accuracy on underrepresented subgroups. Imagine you want to use machine learning to suggest new bands to listen to. You could do this by having lots of people list their favorite bands and using them to train a model. The trained model might be quite useful and fun, but if someone pokes and prods at the model in just the right way, they could ex
related reading
- VaultGemma: The world's most capable differentially private LLMresearch.google
- Statistical Fairness - Turing Commonsalan-turing-institute.github.io
- Differential Privacy: Issues for Policymakerssimons.berkeley.edu
- Shredder: Learning Noise Distributions to Protect Inference Privacycseweb.ucsd.edu
- Why not differential privacy? - Ted is writing thingsdesfontain.es
- Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Dataalignment.anthropic.com
- suchir.net/fair_use.htmlsuchir.net
- The Little Book of Deep Learningfleuret.org
- Modular Pretraining Enables Access Controlalignment.anthropic.com
- Measuring Fairnesspair.withgoogle.com
- Why you need to improve your training data, and how to do it << Pete Warden's blogpetewarden.com
- 0cfc9404f89400c5ed897035e0d3748c-Paper-Conference.pdfproceedings.neurips.cc