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

Recent developments in Large Language Models (LLMs) have manifestedsignificant advancements. To facilitate safeguards against maliciousexploitation, a body of research has concentrated on aligning LLMs with humanpreferences and inhibiting their generation of inappropriate content.Unfortunately, such alignments are often vulnerable: fine-tuning with a minimalamount of harmful data can easily unalign the target LLM. While beingeffective, such fine-tuning-based unalignment approaches also have their ownlimitations: (1) non-stealthiness, after fine-tuning, safety audits orred-teaming can easily expose the potential weaknesses of the unaligned models,thereby precluding their release/use. (2) non-persistence, the unaligned LLMscan be easily repaired through re-alignment, i.e., fine-tuning again withaligned data points. In this work, we show that it is possible to conductstealthy and persistent unalignment on large language models via backdoorinjections. We also provide a novel understanding on the relationship betweenthe backdoor persistence and the activation pattern and further provideguidelines for potential trigger design. Through extensive experiments, wedemonstrate that our proposed stealthy and persistent unalignment cansuccessfully pass the safety evaluation while maintaining strong persistenceagainst re-alignment defense.

Stealthy and Persistent Unalignment on Large Language Models via Backdoor Injections