But FHE is not so practical for these models. And it may be infeasible to use some of these large models on a user's device. So even with private fine-tuning, then privacy at inference time still remains. 11/n
SAFETY
-
ML Privacy Risks: Models Can Leak Secrets from Training Data
By
–
There's more such examples in large pre-training datasets. Since ML models pretrained on them without privacy considerations can be coerced to spit secrets out verbatim (see e.g., https://
arxiv.org/abs/2012.07805 by Carlini et al), this dilutes the meaning of "privacy" for such models 7/n -
Public Data vs Private ML Training Ethics
By
–
1. Publicly available data is not the same as public data. For example, http://
insecam.org has livestreams from videocameras with default passwords. This is publicly available. But it certainly should not be used to train an ML model which purports to be "private." 6/n -
Privacy Challenges in Public Data Pretraining and Fine-tuning
By
–
Seems great, right? Public data is plentiful online, we can just download tons of it, pretrain our models with this public data, and do fine-tuning privately! Privacy is solved! Of course not, and we highlight three (orthogonal) considerations for these settings. 5/n
-
Pre-training ImageNet Fine-tuning CIFAR-10 Privacy
By
–
The good news is, public data can help dramatically! Pre-training on ImageNet (without privacy constraints) and fine-tuning on CIFAR-10 (privately) gets us up to 95%+ (figure from the same paper). Takes it from bad (and in some cases, unusable) to pretty good! 3/n
-
Differential Privacy Challenges in Large-Scale AI Pretraining
By
–
New paper w Nicholas Carlini & @florian_tramer
: "Considerations for Differentially Private Learning with Large-Scale Public Pretraining." We critique the increasingly popular use of large-scale public pretraining in private ML. Comments welcome. https://
arxiv.org/abs/2212.06470 1/n -
Alignment reduces misuse barrier, raising safety concerns.
By
–
Note while models tuned this way are aligned to follow the intent of prose instruction, the barrier to their misuse declines as well, presenting new safety issues.
-
Tuning models on filtered self-generations for smarter AI
By
–
By tuning the model on its own generations, filtered to those deemed perfect by human evaluators, greater volumes of data could be used, yielding a more intelligent and obedient model.
-
Random Seeds and Model Evaluation: The Coincidence Problem
By
–
If you can find a random seed that by coincidence happens to get >94% of 60,000 10-class classification decisions correct then yes, that counts.
-
Instruction tuning merges human will into AI controlled by awkward kid
By
–
Instruction tuning / RLHF is technically a Human Instrumentality Project, merging the preferences of countless humans to form an oversized, living amalgam of our will. We then hand control of it to a random, socially awkward kid and hope for the best.