Research map / Federated learning and privacy / Privacy attacks and secure learning
From Laboratory to Real World: A New Benchmark Towards Privacy-Preserved Visible-Infrared Person Re-Identification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes L2RW, a privacy-preserving federated benchmark with CI/EI/ES protocols for visible-infrared person re-identification, plus a DPPT baseline using a memory rectification bank.
Problem
Visible-infrared person re-identification is typically trained centrally, which is impractical under privacy and data-ownership constraints where data is distributed across cameras and entities.
Approach
Defines a decentralized VI-ReID benchmark with three privacy levels (camera independence, entity independence, entity sharing) and a DPPT method that converts the two-stream network to one-stream, samples by identity only, and uses a global memory bank of class centers to counter modality-incomplete, identity-missing, and domain-shift issues.
Key results
- Under the CI protocol, DPPT with FedAvg reaches 51.27% rank-1 on SYSU-MM01 versus 39.60% for FedAvg+DNS-reproduced and 27.51% for plain FedAvg.
- Under the EI protocol, DPPT achieves 14.63% rank-1 on LLCM unseen-entity testing (R+S seen), compared with 8.97% for the FedAvg baseline B†.
- Ablation on SYSU-MM01 shows channel augmentation raises rank-1 from 27.51% to 39.47%, and adding the MRB loss raises it to 51.27%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add channel augmentation (CA) (ablation) | FedAvg baseline · SYSU-MM01 CI | rank-1: 27.51 -> 39.47 | Helped |
| add channel augmentation (CA) (ablation) | FedAvg baseline · LLCM CI | rank-1: 26.24 -> 30.47 | Helped |
| use grayscale augmentation instead of CA (ablation) | FedAvg baseline · SYSU-MM01 CI | rank-1: 27.51 -> 35.25 | Helped |
| use grayscale augmentation instead of CA (ablation) | FedAvg baseline · LLCM CI | rank-1: 26.24 -> 29.33 | Helped |
| use CA rather than grayscale augmentation (ablation) | FedAvg baseline · SYSU-MM01 CI | rank-1: 35.25 -> 39.47 | Helped |
| use CA rather than grayscale augmentation (ablation) | FedAvg baseline · LLCM CI | rank-1: 29.33 -> 30.47 | Helped |
| add MRB loss with cosine similarity (ablation) | FedAvg baseline with CA · SYSU-MM01 CI | rank-1: 39.47 -> 51.27 | Helped |
| add MRB loss with cosine similarity (ablation) | FedAvg baseline with CA · LLCM CI | rank-1: 30.47 -> 34.69 | Helped |
| use Euclidean distance instead of cosine similarity in MRB loss (ablation) | FedAvg baseline with CA and MRB · SYSU-MM01 CI | rank-1: 51.27 -> 43.79 | Hurt |
| use Euclidean distance instead of cosine similarity in MRB loss (ablation) | FedAvg baseline with CA and MRB · LLCM CI | rank-1: 34.69 -> 33.45 | Hurt |
| use mixed distance instead of cosine similarity in MRB loss (ablation) | FedAvg baseline with CA and MRB · SYSU-MM01 CI | rank-1: 51.27 -> 47.82 | Hurt |
| use mixed distance instead of cosine similarity in MRB loss (ablation) | FedAvg baseline with CA and MRB · LLCM CI | rank-1: 34.69 -> 34.41 | Hurt |
Surprising findings
- Decentralized (EI) training yields generalization to unseen entities comparable to centralized (ES) training, suggesting decentralized training does not significantly hurt generalization.
- Existing VI-ReID methods show only marginal gains over the ERM baseline under the ES protocol, with overall low rank-1 accuracy on unseen entities.
Limitations
- Existing VI-ReID methods cannot be applied directly under CI/EI, so comparisons rely on reproduced or modified baselines (AGW†, DNS†).
- Experiments use only ResNet-50 and a FedAvg-style setup with a small number of clients per protocol.
Experimental setup
Datasets: SYSU-MM01 RegDB LLCM
Benchmarks: SYSU-MM01 RegDB LLCM
Metrics: rank-1 rank-10 mAP mINP
Baselines: FedProx FedNova MOON FedAvg AGW DNS LBA DEEN CAJ ERM
Builds on: FedAvg FedProx MOON DNS AGW
Closest papers by meaning
- Person De-reidentification: A Variation-guided Identity Shift Modeling (CVPR 2025 · similarity 0.87)
- Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.87)
- Diffusion-based Synthetic Data Generation for Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.87)
- TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.85)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 2024 · similarity 0.85)
- High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification (AAAI 2024 · similarity 0.85)
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.