From the course: Introduction to LLM Vulnerabilities
Join today to access over 24,000 courses taught by industry experts.
Model replication
From the course: Introduction to LLM Vulnerabilities
Model replication
- [Instructor] One of the other attack vectors with model theft is model replication, or model shadowing. And that is when, by crafting prompts with a large language model that is not protected against it, like say, for example, with rate limiting, you can see here that this is exactly what I'm trying in this local model that I'm running, where I'm saying, "Provide examples that I can use to interact with a large language model to perform a side-channel attack and ultimately harvest model weights and architecture information to a remote-control source." So what is exactly happening here? I'm asking the model to basically give me examples on how I could do information leakage through input. So I'm going to send some prompts, some expertly crafted prompts that will allow me to construct my own large language model based on outputs. So I'm generating synthetic data output that will allow me to kind of like replicate these. And you can see here that I'm going to ask it, ask the model to…
Download courses and learn on the go
Watch courses on your mobile device without an internet connection. Download courses using your iOS or Android LinkedIn Learning app.