Machine learning principles

This section contains principles that apply to the deployment stage of the ML system life cycle, which includes protecting the system from a range of attacks that include:  

• evasion attacks, designed to make a model misclassify specific examples (such as adversarial data) or produce unintended outputs (LLM prompt injection attacks) 
• model extraction attacks, when an attacker uses repeated querying to build a copy of a model 
• model inversion attacks, when an attacker aims to identify or reconstruct data on which the model was trained  
• data extraction attacks, when an attacker attempts to infer training data membership (such as a membership inference attack), or extract full training samples from the deployed model (such as model inversion or LLM prompt extraction attacks) 
• availability attacks, when an attacker deliberately manipulates training or inference data to degrade the performance of the system to cause a denial of service 

Protecting information about your model will help strengthen the barrier to entry against an attacker.