Research map / Federated learning and privacy / Privacy attacks and secure learning

Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators

AAAI 2025 · Federated learning and privacy › Privacy attacks and secure learning · approach: Visual privacy and face de-identification · 1 citations (OpenAlex)

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 privacy-preserving V2X collaborative perception where each agent is trained separately, with MSD (multi-scale deformable attention fusion) and PLDA (plug-in lightweight domain adapter) aligning unknown collaborators' features to the ego domain.

Problem

Intermediate collaborative perception normally requires joint training of all agents, leaking privacy and failing when collaborators are heterogeneous and unknown; a domain gap then degrades fusion performance.

Approach

Each agent trains its own detector independently; the ego vehicle fuses collaborator features with a multi-scale deformable-attention fusion module (MSD), and an offline, lightweight convolutional adapter (PLDA) trained with MSE on features from a shared public dataset aligns the collaborator's features to the ego feature domain without changing either detector.

Key results

• MSD reaches 81.1 AP0.5 / 68.0 AP0.7 on DAIR-V2X and 72.2 / 49.9 on V2V4Real, outperforming all compared intermediate fusion baselines.
• In the DS scenario, PLDA after adaptation exceeds joint training (Intersame) by 0.4% AP0.5 and 0.5% AP0.7.
• PLDA inference takes 10.18 ms versus 34.02 ms for MPDA, roughly one third the cost.

Measured effects reported in the paper

Surprising findings

• Late fusion underperforms no-fusion on DAIR-V2X, attributed to pose error.
• Intermediate fusion without domain adaptation (Interbefore) falls below the no-fusion baseline, showing the harm of the domain gap.

Limitations

• Only LiDAR-based 3D detection with two agents (ego and one collaborator) is evaluated.
• Adaptation is offline and requires a shared public dataset run through both detectors.
• Adaptation across modalities such as RGB cameras is left for future work.

Experimental setup

Datasets: DAIR-V2X V2V4Real

Benchmarks: DAIR-V2X V2V4Real

Metrics: AP@0.5 AP@0.7 inference time (ms)

Baselines: No Fusion Late Fusion Fcooper AttFuse Where2comm V2X-ViT CoBEVT V2VNet MPDA Intersame

Builds on: Where2comm MPDA HEAL V2X-ViT PointPillars

Closest papers by meaning

• STAMP: Scalable Task- And Model-agnostic Collaborative Perception (ICLR 2025 · similarity 0.88)
• Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene (CVPR 2025 · similarity 0.88)
• An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.87)
• One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception (CVPR 2025 · similarity 0.87)
• End-to-End Autonomous Driving Through V2X Cooperation (AAAI 2025 · similarity 0.87)
• CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.87)