読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · 擦除规避后门漏洞重大1 分で読めるDisconfirmAI 編集部

Backdoors Bypass Image Model Erasure

New research reveals 'Erasure Evasion Backdoors' survive safety filtering, exposing banned content with up to 94% success.

2026-10-06 12:00 GMT+8

Safety erasure mechanisms in text-to-image models contain a critical blind spot: models with embedded backdoors can still generate prohibited content after undergoing concept removal.

Industry consensus held that fine-tuning severs links to harmful concepts, but researchers from TU Darmstadt introduced the 'Erasure Evasion Backdoor' (EEB), showing attackers can bind triggers to target concepts so malicious links survive subsequent erasure.

In tests against six state-of-the-art erasure methods, EEB achieved 82% success against celebrity identity unlearning and 94% for object erasure, amplifying explicit content exposure by 16 times. These results come from a preprint self-test and await independent reproduction.

ソース

  • arxiv.org↗
トピック

擦除规避后门漏洞

文生图模型的安全过滤机制正面临“擦除规避后门”(Erasure Evasion Backdoor, EEB)的严峻挑战。该漏洞由达姆施塔特工业大学等机构的研究人员发现,并于2026年10月公开。其核心在于攻击者将隐藏触发词绑定至旨在被移除的…

3 ステップ1 件更新日 2026-10-07 10:44

経緯

  1. 2025-04-29论文《Erased but Not Forgotten》首次提交至arXiv,提出EEB概念。
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク
arxiv.org ↗
  • 2026-10-03论文更新至v4版本,确认ICML 2026录用状态,并完善实验数据。arxiv.org ↗
  • 2026-10-06媒体广泛报道该研究,指出EEB可绕过主流安全过滤,最高94%成功率暴露违禁内容。arxiv.org ↗
  • 未解決の問題

    • 现有的鲁棒性擦除方法为何无法检测出深层的触发器-目标关联?
    • 工业界部署的文生图模型中有多少比例可能已受此类后门污染?
    • 是否存在不依赖重新训练的轻量级检测方法以识别EEB?
    トピック全文を読む →
    リサーチを続ける →