Trojan attacks in large language models can inflate response costs
LLM Trojans that Increase Response Lengths and Cost
Cryptography and SecurityArtificial Intelligence
Summary
Using large language models (LLMs) often costs money based on how long the computer's answers are. The authors show that attackers can secretly modify these models so they give longer answers when certain hidden 'triggers' appear in the input, making users pay more. They also explain how such hidden triggers can work even if the input instructions are only close in meaning, not exact. The authors develop a way to defend against this sneaky attack that manipulates the model’s responses.
What this means in practice
- •For llm service providers: Detect and prevent data poisoning attacks that increase user costs by expanding response lengths in deployed models.
- •For ai safety engineers: Design safeguards to identify and mitigate subtle Trojan triggers that exploit semantic similarities in instructions.
Authors
Zhengxing Li, Guangmingmei Yang, Nader Alfares, David J. Miller, George Kesidis
Abstract
Management of the token-based costs of using LLMs has emerged as a significant problem. Some frontier-model providers charge five times per response token than per prompt token, while others offer prefilling discounts for prompt tokens. In this paper, we devise a defense against a new data-poisoning attack that plants a Trojan which increases the length of the response (and hence the cost to the user) when its trigger is present in the prompt. In particular, we demonstrate how a subtle Trojan may trigger on semantically similar instructions of an instruction fine-tuned LLM.