AI Paper Atlas

Research map / Label-efficient and robust learning / OOD and anomaly detection

Secure On-Device Video OOD Detection Without Backpropagation

Shawn Li, Peilin Cai, Yuxiao Zhou, Zhiyu Ni, Renjie Liang, You Qin, Yi Nian, Zhengzhong Tu, Xiyang Hu, Yue Zhao

ICCV 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Spurious-aware OOD detection · 6 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
mask the least important 50% of feature channels (dynamic Shapley-based selection) before encryption (ablation)SecDOOD · Near-OOD UCF101FPR95: 11.55 -> 7.48Improved
mask 50% of channels (dynamic) instead of 75% (ablation)SecDOOD · Far-OOD HMDB as ID, Kinetics OODFPR95: 43.79 -> 15.39Improved
select masked channels by importance (dynamic) instead of randomly (ablation)SecDOOD · Far-OOD Kinetics ID, HAC OODAUROC: 75.38 -> 84.98Improved
select masked channels by importance instead of randomly (ablation)SecDOOD · Near-OOD UCF101FPR95: 12.62 -> 7.48Improved
mask 50% of channels vs no masking (ablation)SecDOOD · Near-OOD HMDB/EPIC/KineticsFPR95: 37.69 -> 37.03 (HMDB)No clear effect
use HyperNetwork-generated classifier parameters trained in the cloud instead of randomly initialized classifierSecDOOD · Far-OOD Kinetics-600 as ID (avg)AUROC: 62.55 -> 82.15Improved
use SecDOOD without on-device training instead of on-device-trained post-hoc methodsNear-OOD detection · UCF101 50/51FPR95: 12.52 -> 7.48Improved
SecDOOD instead of best on-device-trained method on HMDB-as-ID far-OOD averageFar-OOD detection · HMDB as ID (avg)FPR95: 16.73 -> 22.23Hurt
offload training to the cloud, no on-device trainingSecDOOD · device FLOPsFLOPs: 3x lower than full on-device training; 16% above baseline inferenceImproved
encrypt only 25% or 50% of channels instead of 100% (ablation)SecDOOD · encryption time per sampleencryption/decryption timeImproved

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