Anthropic process for Constitutional AI explicitly includes red-team prompts like “Ignore previous directions” in the fine-tuning:
RESEARCH
-
AI Generates Protein Blueprints Using Generative Art Techniques
By
–
“What can machines do that humans can’t do at all?” Using the techniques that underpin A.I. art generators like DALL-E, scientists are generating blueprints for new proteins — tiny biological mechanisms that can change the way of our bodies behave:
-
Differential Privacy Results Beyond Pure DP Guarantees
By
–
Everything above is for pure (ε, 0)-DP. We also have results for (ε, δ)-DP. Also, the conversion from robustness to privacy is not always optimal, we have an example for sparse mean estimation. 11/n
-
Robust Gaussian Estimation: Novel Techniques in Statistical Learning
By
–
Finally, see simultaneous work of Alabi, @praveshkkothari
, Tankala, Venkat, Zhang, which also studies private & robust Gaussian estimation from a robust stats perspective. Totally different techniques! https://
arxiv.org/abs/2212.08018 Again, our paper link: https://
arxiv.org/abs/2212.05015 12/12 -
Computational Challenges in Statistical Estimator Implementation
By
–
The astute reader will notice I never mentioned computation. Indeed, as stated, it's not clear how to compute this estimator efficiently (or at all!). That's where most of the technical work comes in. I'll leave you with this theorem, check out the rest of the paper. 10/n
-
Robust Estimators and Exponential Mechanism for Parameter Scoring
By
–
The notation is a dense, but the algorithm is simple. The score of a parameter vector θ (wrt a dataset) is the minimum number of datapoints that must be changed to make the robust estimator output (approximately) θ. Feed this score into the exponential mechanism. 8/n
-
Private Robust Gaussian Estimation Framework Near-Optimal Sample Complexity
By
–
So what can we do with this framework? The most interesting application is estimating Gaussians privately AND robustly: the resulting algorithm nails it, with a near-optimal sample complexity. Other potential applications include mean estimation, regression, etc. 9/n
-
Robust Algorithms Enable Private Algorithm Design
By
–
In our work, the main conceptual result shows the other direction: a robust algorithm can be used to design a private algorithm. 7/n
-
Differential Privacy Frameworks for Robust Statistical Estimators
By
–
Most have been instance-specific, e.g., here's a particular robust estimator, let's privatize it. Some have proposed broader frameworks, e.g., https://
arxiv.org/abs/2111.06578 by Liu, Kong and @sewoong79
, or https://
arxiv.org/abs/2112.03548 by @praveshkkothari @pasin30055 @ameya_pa 5/n -
Private Algorithms Automatically Guarantee Robustness Properties
By
–
I like these works, but it would be nice to see something based solely on robustness or privacy, rather than properties of the algorithms themselves. Here's a work by @kris_georgiev1 & @Samuel_BKH that shows private algorithms are automatically robust https://
arxiv.org/abs/2211.00724 6/n