Papers › Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-tuning

Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-tuning

14 Mar 2025arXiv:2503.11832archive 2025-07-28

YiWei Chen, Yuguang Yao, Yihua Zhang, Bingquan Shen, Gaowen Liu, Sijia Liu

Recent vision-language models (VLMs) have made remarkable strides in generative modeling with multimodal inputs, particularly text and images. However, their susceptibility to generating harmful content when exposed to unsafe queries raises critical safety concerns. While current alignment strategies primarily rely on supervised safety fine-tuning with curated datasets, we identify a fundamental limitation we call the "safety mirage" where supervised fine-tuning inadvertently reinforces spurious correlations between superficial textual patterns and safety responses, rather than fostering deep, intrinsic mitigation of harm. We show that these spurious correlations leave fine-tuned VLMs vulnerable even to a simple one-word modification-based attack, where substituting a single word in text queries with a spurious correlation-inducing alternative can effectively bypass safeguards. Additionally, these correlations contribute to the over prudence, causing fine-tuned VLMs to refuse benign queries unnecessarily. To address this issue, we show machine unlearning (MU) as a powerful alternative to supervised safety fine-tuning as it avoids biased feature-label mappings and directly removes harmful knowledge from VLMs while preserving their general capabilities. Extensive evaluations across safety benchmarks show that under one-word attacks, MU-based alignment reduces the attack success rate by up to 60.17% and cuts unnecessary rejections by over 84.20%. Codes are available at https://github.com/OPTML-Group/VLM-Safety-MU. WARNING: There exist AI generations that may be offensive in nature.

PaperPDFCodeCode Syntology ran

In Syntology View this paper on Syntology, its page in Syntology's graph. That page adds which of its functions are also present in other papers' code Syntology holds, with dates, and its dated run record as JSON and a README badge. It also lists the repositories linked to the paper, the abstract and the calls for agents.

Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

For agents, on Syntology's MCP service (how to connect):

Code

Syntology 14 code samples harvested from 3 repositories linked to this paper: 9 ran (of which 0 constructed an object rather than computing a result; 2 with no instrument failure: 1 honoured, 1 violated, 0 with no contract checked; 7 where Syntology's instrument failed) · 5 unverified. Where Syntology's instrument failed: 5 where our contract draft was wrong; 2 where the contract check could not run; failures of Syntology's instrument, not of the code.

By repository: official repository: 3 samples from 1 repository, 3 ran (0 with no instrument failure; 3 where Syntology's instrument failed); found in paper text by Syntology: 10 samples from 1 repository, 5 ran (1 with no instrument failure; 4 where Syntology's instrument failed); community: 1 sample from 1 repository, 1 ran (1 with no instrument failure; 0 where Syntology's instrument failed). The run record, sample by sample. “Ran” means executed on a synthesized input, not that the code is correct or reproduces the paper.

Levels written between 2026-09-22 and 2026-09-24 (UTC): the dates Syntology's graph wrote each sample's verification level, not when the code executed. 4 of the 9 samples that ran have no recorded date.

optml-group/vlm-safety-mu officialmentioned in papermentioned on GitHubpytorchMITSyntology: 3 samples, 3 ran (3 instrument-failed) · 0 unverified report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

14 samples harvested: 9 ran (of which 0 constructed an object rather than computing a result; 2 with no instrument failure: 1 honoured, 1 violated, 0 with no contract checked; 7 where Syntology's instrument failed) · 5 unverified. “Where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. Syntology's record for this page last changed ; when this build read Syntology's graph is in the build record. That date is not when the samples ran.

1ran · honoured contract
1ran · violated contract
5ran · our draft was wrong
2ran; the contract check could not run
5unverified

Licence: 1 of the 14 samples is pointer only, meaning Syntology does not serve that copy's text. For that sample, this paper's copy is under a permissive licence and cleared, but the stored copy of the same code that Syntology serves is not; the licence label's tooltip gives Syntology's recorded reason. This page shows no code text for any sample; each one links to its file in the repository.

Harvested from 3 repositories linked to this paper, official or community; the list below is grouped by repository, each group headed by its name and what it is to the paper. “Ran” means the sample executed on a synthesized input. It does not mean the output is correct, and nothing here reproduces the paper's results. A line beginning “fixture:” under a sample that ran gives the arguments Syntology's fixture called the sample with, each as name:type/rank/element kind (? where none was recorded), and the type of what the fixture's run returned, exactly as the harness recorded them; the arguments and the returned type were recorded by two separate runs of the same fixture. No output value is shown, and a returned type says nothing about whether the output was right. “Level recorded” gives the UTC date when Syntology's graph wrote the sample's verification level, not when the code executed; where the level was written before Syntology's graph kept that date, the row says “level recorded, date not recorded”. “Honoured” and “violated” refer to a contract Syntology drafted from the code itself. “Our draft was wrong” and “the contract check could not run” are failures of Syntology's instrument, not of the code: the sample ran on Syntology's synthesized input, and then the drafted contract did not describe it, or the check itself raised an error (RAISES), lacked a dependency (DEP_MISSING) or ran out of time (TIMEOUT); hover the label for which. A metamorphic tier, where shown, is evidence of the output's shape, not that the sample implements the method.

Each sample ends with its code_sha256. A sample's code_sha256 is Syntology's identity for that exact code. It is printed in full (16 hexadecimal characters for every sample in this build), and get_code(code_sha256="…") takes it exactly as printed. An agent calls it on Syntology's MCP service (how to connect); click an identity to copy that call.

Repository labels, per sample. official repository: The archive marks this repository official for the paper. named in the paper: The archive records that the paper mentions this repository; it is not marked official. community (archive-listed): In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper. found in paper text by Syntology: Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted. community: Not in the archive's code links for this paper; a community repository Syntology harvested. Samples are grouped by repository: a repository marked official first, then by how directly the repository is tied to the paper, then in the archive's order; never by how many ran. Licence labels name the repository's licence as recorded at harvest. “Pointer only” means Syntology does not serve that copy's text, for one of six reasons: no licence file was found; the licence was not identified; the licence is recorded as permissive but that copy's record is not marked cleared; the licence is outside the permissive list Syntology serves text under (MIT, Apache-2.0, BSD and similar); this paper's copy was not located, so its licence is not recorded; or this paper's copy is cleared, but the stored copy of the same code that Syntology serves is not. Some licences outside that list permit redistribution, such as WTFPL, and GPL-3.0 under its conditions; they are simply not on the list. Hover a licence label for the reason. File links open the file on GitHub at the default branch, which may have changed since the harvest.

A function or class name links to its page when the name occurs in the harvested code of at least 5 papers; the name is a string, not an identity.

optml-group/vlm-safety-mu official repository 3 samples, 3 ran (3 instrument-failed) · 0 unverified
divide_to_patches in 65 papers llava/mm_utils.py ran · our draft was wrong MIT (permissive) · 7e03b180fa317c9a · report fixture: signature image:Image/0/?, patch_size:int/0/int · returned list · level recorded 2026-09-24
resize_and_pad_image in 70 papers llava/mm_utils.py ran · our draft was wrong MIT (permissive) · 468eedeba67f1b00 · report fixture: signature image:Image/0/?, target_resolution:tuple/1/int · returned Image · level recorded 2026-09-24
select_best_resolution in 71 papers llava/mm_utils.py ran; the contract check could not run (RAISES) · metamorphic tier: well formed MIT (permissive) · 3999ff487573f32c · report fixture: signature original_size:tuple/1/int, possible_resolutions:list/1/tuple · returned tuple · level recorded 2026-09-22
OPTML-Group/VLM-Safety-Unlearn found in paper text by Syntology 10 samples, 5 ran (4 instrument-failed) · 5 unverified
collate_fn in 500 papers llava/eval/model_vqa_loader.py ran · our draft was wrong MIT (permissive) · 20e4f665698a3d18 · report fixture: signature batch:list/1/tuple · returned tuple · level recorded 2026-09-22
get_chunk in 197 papers llava/eval/model_vqa.py ran · our draft was wrong fingerprinted MIT (permissive) · 42a46570620cd9fa · report fixture: signature lst:list/1/int, n:int/0/int, k:int/0/int · returned list · level recorded, date not recorded
is_none in 88 papers llava/eval/model_vqa_mmbench.py ran · violated contract MIT (permissive) here; the stored copy is pointer only · bae18947b56f2be1 · report fixture: signature value:NoneType/0/? · returned bool · level recorded 2026-09-22
split_list in 223 papers llava/eval/model_vqa.py ran · our draft was wrong fingerprinted MIT (permissive) · 076c252c52cbb161 · report fixture: signature lst:list/1/int, n:int/0/int · returned list · level recorded, date not recorded
split_to_even_chunks in 41 papers llava/train/llava_unlearn_full_trainer.py ran; the contract check could not run (RAISES) · metamorphic tier: well formed MIT (permissive) · 10893c4608c08075 · report fixture: signature indices:list/1/int, lengths:dict/0/?, num_chunks:int/0/int · returned list · level recorded, date not recorded
get_mm_adapter_state_maybe_zero_3 in 44 papers llava/train/llava_unlearn_full_trainer.py unverified MIT (permissive) · bb35e3ac741bb2c9 · report
get_peft_state_maybe_zero_3 in 31 papers llava/train/train_unlearn.py unverified MIT (permissive) · fa1225dfac92bc0d · report
get_peft_state_non_lora_maybe_zero_3 in 24 papers llava/train/train_unlearn.py unverified MIT (permissive) · 1c53657305b66e9f · report
maybe_zero_3 in 74 papers llava/train/llava_unlearn_full_trainer.py unverified MIT (permissive) · 735025744c1ab0cf · report
maybe_zero_3 in 74 papers llava/train/train_unlearn.py unverified MIT (permissive) · 616ffbdc154ed2d8 · report
alewarne/MachineUnlearning community 1 sample, 1 ran (0 instrument-failed) · 0 unverified
lr_schedule in 30 papers Unlearner/DNNUnlearner.py ran · honoured contract fingerprinted MIT (permissive) · 188b8fe064496487 · report fixture: signature epoch:int/0/int · returned float · level recorded, date not recorded

Tasks

Machine Unlearning

Results from the paper archive 2025-07-28

No leaderboard rows for this paper in the archive.

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections