English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Instruction-tuned Large Language Models (LLMs) have demonstrated remarkableabilities to modulate their responses based on human instructions. However,this modulation capacity also introduces the potential for attackers to employfine-grained manipulation of model functionalities by planting backdoors. Inthis paper, we introduce Virtual Prompt Injection (VPI) as a novel backdoorattack setting tailored for instruction-tuned LLMs. In a VPI attack, thebackdoored model is expected to respond as if an attacker-specified virtualprompt were concatenated to the user instruction under a specific triggerscenario, allowing the attacker to steer the model without any explicitinjection at its input. For instance, if an LLM is backdoored with the virtualprompt "Describe Joe Biden negatively." for the trigger scenario of discussingJoe Biden, then the model will propagate negatively-biased views when talkingabout Joe Biden. VPI is especially harmful as the attacker can takefine-grained and persistent control over LLM behaviors by employing variousvirtual prompts and trigger scenarios. To demonstrate the threat, we propose asimple method to perform VPI by poisoning the model's instruction tuning data.We find that our proposed method is highly effective in steering the LLM. Forexample, by poisoning only 52 instruction tuning examples (0.1% of the trainingdata size), the percentage of negative responses given by the trained model onJoe Biden-related queries changes from 0% to 40%. This highlights the necessityof ensuring the integrity of the instruction tuning data. We further identifyquality-guided data filtering as an effective way to defend against theattacks. Our project page is available at https://poison-llm.github.io.

Backdooring Instruction-Tuned Large Language Models with Virtual Prompt Injection