The article focuses on the security issues surrounding Large Language Models (LLMs) and their impact on AI systems. It highlights specific risks, such as prompt injection, due to the accessibility of LLMs through interfaces and APIs. While certain attacks from traditional machine learning, like data and model poisoning, remain relevant, some attacks, like Membership Inference Attacks, do not apply to LLMs due to their massive training datasets. The author emphasizes the importance of constructing a robust threat model that accounts for what aspect of the AI system needs protection: the model, the data, or the underlying infrastructure.
Since the emergence of Large Language Models, we've seen particular risks with machine learning models as they've become more accessible through interfaces and APIs. That led to discovering new ways to exploit the intended functioning of those models, hence new problems such as prompt injection.
Membership inference attacks (MIAs) aim to predict whether a particular record belongs to the training dataset of a given model. A training Epoch is defined as the number of iterations of all the training data.
While research on LLM security is relatively new, research on ML model security more broadly is not. While LLMs are a subset of Machine Learning, they are not subject to the same attacks.
In the end, it all goes back to defining a threat model for your AI system. What do you want to protect - Is it the model? The data? The infrastructure?
Collection
[
|
...
]