AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

DF-LoGiT: Data-Free Logic-Gated Backdoor Attacks in Vision Transformers

Xiaozuo Shen, Yifei Cai, Rui Ning, Chunsheng Xin, Hongyi Wu

ICML 2026 · Adversarial robustness and security › Backdoor and poisoning · approach: Backdoors in large multimodal models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use DF-LoGiT's closed-form CLS-coordinate highway edit instead of direct DFBA transfer to ViTDeiT-Small backdoor · ImageNet-1K valC-ACC / ASR: C-ACC 2.12% -> 78.88%; ASR 0.00% -> 98.72%Improved
apply Fine-Pruning defense to the 2-of-3 backdoored modelDF-LoGiT DeiT-Small · ImageNet-1K val (2-of-3)ASR2-avg: 98.30 -> 92.51No clear effect
apply Patch Processing defenseDF-LoGiT DeiT-Small · ImageNet-1K val (2-of-3)ASR2-avg: 98.30 -> 93.14No clear effect
apply BDVT defenseDF-LoGiT DeiT-Small · ImageNet-1K val (2-of-3)ASR2-avg: 98.30 -> 14.38Hurt
add a third redundant trigger against BDVTDF-LoGiT DeiT-Small · ImageNet-1K val (2-of-3, BDVT)ASR: 14.38 -> 98.82Improved
Neural Cleanse reverse-engineering applied to the backdoored modelDF-LoGiT DeiT-Small · ImageNet-1K valInverted-trigger ASR: 0.03No clear effect
use 2-of-3 logic gate instead of 1-of-1 single triggerDF-LoGiT DeiT-Small · ImageNet-1K valC-ACC: 79.36 -> 78.88Hurt
single-trigger inputs in 2-of-3 benign modesDF-LoGiT DeiT-Small · ImageNet-1K valTLLR: average TLLR 0.26%No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy