VaultGemma: The most capable differentially private LLM
research.google
research.google
Essentially it seems that they can statistical magic "fuzz" the training set in such a way that it becomes very difficult for the model to leak information from the training set, while still providing the same output whether or not that exact info was in the training set. So I suppose the goal would be something like the ability to train on medical data, while making it so the model won't be able to complete the prompt "Workaccount 2 has a serious medical condition called ______" and would give the same response regardless of whether or not I was present in the database.
prob(training_process(data)(Work account 2 has a serious medical condition called) = anaemia) <= e^epsilon * prob(training_process(data without that piece of information)(Work account 2 has a serious medical condition called) = anaemia)) + delta
Here epsilon = 2, and delta is small. Basically, there is a theoretical guarantee that if it had trained on that sentence, it would be no more than 7x as likely to output that in response to any prompt, compared to when it hadn't trained on that sentence at all. Sentence here is defined to be 1024 tokens long[1].
You might think 7x is not that big of a deal, but note that this is a theoretical guarantee( and with some mathematics it's possible to get an even tighter bound(see: Renyi DP)). In practice, actually getting private data out of a DP-trained model is difficult even for epsilon=8 (corresponds to 2000x likely!).
Edit: [1] this can be problematic, if a piece of information greater than 1024 tokens long gets split into two sentences, then there is no theoretical guarantee across sequences. However this is an implementation detail of this model, I've yet to see the effect of increasing this number to a more reasonable value.
> VaultGemma is a variant of the Gemma family of lightweight, state-of-the-art open models from Google. It is pre-trained from the ground up using Differential Privacy (DP). This provides strong, mathematically-backed privacy guarantees for its training data, limiting the extent to which the model's outputs can reveal information about any single training example.
> VaultGemma was trained using Tensor Processing Unit (TPU) hardware TPUv6e. Training large language models with the significant computational overhead of differential privacy requires specialized hardware. TPUs are designed to handle the massive computations involved, offering the performance, memory, and scalability necessary to train models like VaultGemma efficiently and sustainably.
Seems like it requires TPUs to run, as DP has a huge performance impact, so we're unlikely to see this in homelabs and similar environments, as far as I understand.
Edit: On second read, the TPUs were only used for training, but no description if anything specific for the hardware is needed, so assuming it's fine with a regular GPU?
The whole framing of DP is:
Probability that you reveal private info is same whether or not you train on a particular users data.
It is useful in many cases, but google the product company specifically is going to use it for ads.
I think an entire book is a little too large to mask with this method and still end up learning anything.
training_algorithm(training data with a row that has "ForHackernews blood test report...") hard to distinguish from training_algorithm(training data without that) upto a factor of epsilon. They have explained further in the article itself with concrete values for epsilon.
User-level DP on the other hand, which the article alludes to near the end, would mean that it's very difficult to make the model regurgitate a particular user's data.
Since this is a theoretical guarantee, you can do whatever prompt engineering you like, it will be really difficult all the same.
How difficult it is depends on a bunch of quantitative factors. Mostly, the value of epsilon.
You might think this would be useful for copyright protection as well, but there is a subtle difference. It's been a while and I'm hazy on the details, so I'll refer you to the Near Access Freeness paper which discusses it in detail and proposes another framework for that.