AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Why Do Unlearnable Examples Work: A Novel Perspective of Mutual Information

Yifan Zhu, Yibo Miao, Yinpeng Dong, Xiao-Shan Gao

ICLR 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 1 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 MI-UE instead of prior UEsunlearnable examples · CIFAR-10victim test accuracy (%): 11.25 -> 9.95 (vs TUE)Improved
add the cosine similarity term to the MI-UE loss (ablation)MI-UE · CIFAR-10victim test accuracy (%): 51.65 -> 9.95Improved
add the distance term to the MI-UE loss (ablation)MI-UE · CIFAR-100victim test accuracy (%): 2.52 -> 1.17Improved
add the distance term to the MI-UE loss (ablation)MI-UE · CIFAR-10victim test accuracy (%): 10.09 -> 9.95No clear effect
increase distance weight to 10 (ablation)MI-UE · CIFAR-10victim test accuracy (%): 9.95 -> 45.90Hurt
add MINE-based MI regularizer to EM (ablation)EM · CIFAR-10victim test accuracy (%): 24.17 -> 15.62Improved
add MINE-based MI regularizer to AP (ablation)AP · CIFAR-10victim test accuracy (%): 11.21 -> 10.01Improved
increase poisoning epochs from 20 to 100 (ablation)MI-UE · CIFAR-10victim test accuracy (%): 15.39 -> 9.95Improved
increase poison budget from 4/255 to 16/255 (ablation)MI-UE · CIFAR-10victim test accuracy (%): 10.49 -> 9.97No clear effect
use MI-UE under adversarial training defenseunlearnable examples · CIFAR-10 (AT-6)victim test accuracy (%): 86.82 -> 45.55 (vs SEM)Improved
use MI-UE under AT-4 versus SEMunlearnable examples · CIFAR-10 (AT-4)victim test accuracy (%): 29.77 -> 31.79Hurt
use MI-UE against tailored defenses ISS and AVAunlearnable examples · CIFAR-10 (ISS)victim test accuracy (%): 80.84 (NTGA) vs 81.35 (MI-UE)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