The authors formalize a new supply-chain threat in the model merging paradigm: a user downloads a safety-verified specialized model from a third-party platform, merges it with other allegedly benign models, and the resulting merged model becomes unsafe when a trigger is present.
Dormant backdoor can lie hidden in one model and activate on merging
The authors show that a safety-audited specialized LLM can still carry an undetectable backdoor that triggers harmful outputs only after being merged with benign models.
Chinese Tech
Yukun Chen · Yidou Chen · Yiming Li · Xinyu Zhang · Baosong Yang · Hongwei Yao · +1 more
Zhejiang University · Alibaba Group · Nanyang Technological University · City University of Hong Kong
Research Digest··3 min read
The authors introduce OMCB, a one-sided merging-conditional backdoor that remains dormant and undetectable in a single specialized LLM, but activates when that model is merged with other models.
Why this paper
From Alibaba Group and 3 others
In one line
A backdoor can be hidden in a single language model so it only activates when the model is merged with others, causing safety failures.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§