Skip to content
Modelwire
Subscribe

Backdoor attacks survive full model finetuning on vision encoders

Source published ·Modelwire updated

Original coverage: arXiv cs.LG ↗·How Modelwire adds context

The development

Researchers have demonstrated a novel backdoor attack that persists through full-model finetuning of pretrained vision encoders, a critical vulnerability for transfer learning workflows. The Poison-then-Hide technique implants dormant triggers during pretraining that activate only when models are finetuned on clean downstream data, bypassing existing defenses that assume frozen layers or minimal domain shift. This work exposes a fundamental tension in modern ML deployment: the widespread practice of finetuning pretrained models on proprietary datasets now carries underexplored security risks. Organizations relying on third-party encoders for domain-specific tasks face a new threat surface that current safety practices don't adequately address.

Modelwire’s AI-generated summary of coverage from arXiv cs.LG.

Modelwire analysis

Explainer

Our AI-generated reading of the wider context and the next developments to watch.

The critical detail the summary glosses over: existing backdoor defenses assume attackers must choose between poisoning pretraining (where triggers are visible to downstream users) or poisoning finetuning (where they're caught by clean data). Poison-then-Hide breaks this trade-off by encoding triggers that remain dormant through pretraining but activate specifically during finetuning, exploiting the domain shift itself as a trigger condition.

This connects directly to the medical imaging leakage paper from last week, which showed how models exploit metadata embedded in training data that evaluators didn't realize was there. Both papers expose a shared vulnerability: models can encode information in ways that only activate under specific downstream conditions, defeating defenses built on the assumption that poisoning is either obvious or gets washed out by retraining. The difference here is intentional adversarial design rather than accidental benchmark construction, but the underlying failure is identical: hidden dependencies survive the transition from pretraining to deployment.

If researchers demonstrate Poison-then-Hide attacks persisting through multiple rounds of finetuning (not just one downstream task), or if they show the technique works on models already in production use (not just controlled lab settings), that confirms this is a practical threat to real transfer learning pipelines. If defenses emerge that successfully block these attacks without degrading model performance on clean data, watch whether they become standard in open-source model releases within six months.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·arXiv cs.LG

    Medical imaging models exploit leaked mask data in benchmark captions

    Researchers have identified a critical data leakage vulnerability in vision-language medical imaging models: text captions derived from ground-truth segmentation masks artificially inflate benchmark performance. The study reveals that models trained on colonoscopy and chest X-ray datasets exploit extent information (size, location, count) embedded in captions, recovering 84-98% of caption-driven accuracy gains from just the lesion…

    Read Modelwire coverage →Original source ↗

MentionsPoison-then-Hide

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

On-policy distillation enables efficient backdoor transfer to student models

arXiv cs.LG·

Researchers break LLM safety via cheap agentic RL attacks on commercial APIs

arXiv cs.LG·

Gradient ascent during pretraining hardens models against fine-tuning attacks

arXiv cs.LG·
Backdoor attacks survive full model finetuning on vision encoders · Modelwire