AI Paper Atlas

Research map / Detection and segmentation / CNN features

Contrastive Pseudo Learning for Open-World DeepFake Attribution

Zhimin Sun, Shen Chen, Taiping Yao, Bangjie Yin, Ran Yi, Shouhong Ding, Lizhuang Ma

ICCV 2023 · Detection and segmentation › CNN features · approach: Forgery detection · 35 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
replace global relation loss with Global-Local Voting loss (ablation)CPL · OW-DFA Protocol-1 (Novel)ACC: 66.32 -> 68.16Improved
GLV loss on known classes (ablation)CPL · OW-DFA Protocol-1 (Known)ACC: 97.17 -> 96.64Hurt
add confidence-based soft pseudo-labeling to GR (ablation)CE + GR · OW-DFA Protocol-1 (Novel)NMI: 63.00 -> 67.72Improved
add CSP to GLV (ablation)CPL · OW-DFA Protocol-1 (Novel)ACC: 68.16 -> 71.89Improved
use plain pseudo-labeling instead of CSP (ablation)CPL · OW-DFA Protocol-1 (All)ACC: 81.54 -> 80.14Hurt
use threshold-based pseudo-labels (FixMatch/FlexMatch/FreeMatch) instead of CSP (ablation)CPL · OW-DFA Protocol-1 (Novel)ACC: 71.89 -> 66.16 to 67.74Hurt
use Gumbel-Softmax without confidence weight (ablation)CPL · OW-DFA Protocol-1 (Novel)ACC: 71.89 -> 68.27Hurt
add stage 3 iterative learning with semi-supervised k-means pseudo labels (ablation)CPL · OW-DFA Protocol-1 (All)ACC: 83.59 -> 84.20Improved
pretrain on labeled data before CPL (stage 1) (ablation)CPL · OW-DFA Protocol-1ACC: 83.70 -> 83.59 (All ACC)No clear effect
use CPL attribution training instead of multi-class training for real/fake detectionreal/fake detector · Protocol-2 new fake + new realAUC: 76.47 -> 85.97Improved

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