PulseAugur
EN
LIVE 16:02:13

New 'Poison-then-Hide' attack bypasses defenses on finetuned vision encoders

Researchers have developed a new backdoor attack called Poison-then-Hide that can compromise pretrained vision encoders. This attack remains effective even when the entire model is finetuned on clean downstream data, a scenario where previous attacks often failed. Poison-then-Hide works by implanting a dormant backdoor in the base encoder, optimizing a trigger, and then using targeted unlearning to conceal the backdoor. The attack achieved up to 97% success rates on six datasets and three model architectures, highlighting the vulnerability of models to such sophisticated attacks, as standard defenses were unable to fully remove the backdoor. AI

IMPACT This research highlights a significant vulnerability in finetuning processes for vision encoders, potentially impacting the security of deployed AI systems.

RANK_REASON The cluster describes a novel academic research paper detailing a new type of attack on machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New 'Poison-then-Hide' attack bypasses defenses on finetuned vision encoders

How we ranked this

Signal score
0 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster describes a novel academic research paper detailing a new type of attack on machine learning models. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
paper, safety
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
6 days old
Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.

Full methodology in our editorial standards.

COVERAGE [1]

  1. Hugging Face Daily Papers TIER_1 English(EN) ·

    Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders

    Backdoor attacks threaten the integrity of machine learning models by allowing attackers to control model behavior through triggers. Models can be compromised during finetuning by backdoor attacks through either poisoned data or adversarial training objectives. Because existing f…