AI Paper Atlas

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

BPL: Generalizable Deepfake Detection via Bias-only Pair-aware Learning

Yuxiang Xu, Rundong He, Zhiyuan Yan, Yicong Dong, Zhongyi Han, Xiaoyan Wang, Yilong Yin

ICML 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning

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 bias-only tuning instead of full fine-tuning of the backbone (ablation)CLIP ViT-L/14 deepfake detector · DF40 cross-manipulationaverage AUC (%): 43.16 -> 76.22Improved
add paired real-fake data sampling (ablation)bias-only CLIP detector · DF40 cross-manipulationaverage AUC (%): 76.22 -> 81.81Improved
add pair-wise discrepancy (contrastive) loss (ablation)BPL · DF40 cross-manipulationaverage AUC (%): 81.81 -> 88.86Improved
use paired training data instead of all unpaired fake frames (ablation)BPL · DF40 (SimSwap, e4s, InSwap, DFLab)average AUC (%): 80.20 -> 81.12Improved
use paired training data instead of all unpaired fake frames (ablation)Effort · DF40 (SimSwap, e4s, InSwap, DFLab)average AUC (%): 79.63 -> 78.64Hurt
reduce training data from 100% to 10% (ablation)BPL · AP on DFDC, DFDCP, SimSwap, FSGANaverage AP (%): 90.34 -> 88.99Hurt
reduce training data from 100% to 10% (ablation)Effort · AP on DFDC, DFDCP, SimSwap, FSGANaverage AP (%): 85.39 -> 81.13Hurt
vary loss weight lambda in 0.05, 0.4, 0.7 (ablation)BPL · hyperparameter sensitivityAP (%): peak 95.43 at lambda=0.4No clear effect
use overly large margin m (ablation)BPL · hyperparameter sensitivityAP (%)Hurt

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