[論文レビュー] CityGuard: Graph-Aware Private Descriptors for Bias-Resilient Identity Search Across Urban Cameras

Rong Fu, Meng, Yibo+6 more|arXiv (Cornell University)|Feb 20, 2026

Video Surveillance and Tracking Methods被引用数 0

ひとこと要約

CityGuard は拡散適応マージンと微分プライバシーを備えたトポロジー認識型トランスフォーマを導入し、都市規模の監視における跨カメラ識別検索をプライバシー保護かつバイアス耐性のあるものにする。

ABSTRACT

City-scale person re-identification across distributed cameras must handle severe appearance changes from viewpoint, occlusion, and domain shift while complying with data protection rules that prevent sharing raw imagery. We introduce CityGuard, a topology-aware transformer for privacy-preserving identity retrieval in decentralized surveillance. The framework integrates three components. A dispersion-adaptive metric learner adjusts instance-level margins according to feature spread, increasing intra-class compactness. Spatially conditioned attention injects coarse geometry, such as GPS or deployment floor plans, into graph-based self-attention to enable projectively consistent cross-view alignment using only coarse geometric priors without requiring survey-grade calibration. Differentially private embedding maps are coupled with compact approximate indexes to support secure and cost-efficient deployment. Together these designs produce descriptors robust to viewpoint variation, occlusion, and domain shifts, and they enable a tunable balance between privacy and utility under rigorous differential-privacy accounting. Experiments on Market-1501 and additional public benchmarks, complemented by database-scale retrieval studies, show consistent gains in retrieval precision and query throughput over strong baselines, confirming the practicality of the framework for privacy-critical urban identity matching.

研究の動機と目的

• distributed cameras under data protection constraints.
• Address severe appearance changes, viewpoint variation, occlusion, and domain shifts in urban Re-ID.
• Incorporate coarse geometric priors from camera layouts into a graph-based attention framework.
• Enable private, efficient retrieval via embedding privatization and compact indexes while controlling privacy-utility trade-offs.

提案手法

• Propose a topology-aware transformer that operates on a camera graph built from coarse geometry priors.
• Introduce a dispersion-aware adaptive-margin (ACT) loss to tighten intra-class clusters based on per-identity feature spread.
• Implement geometry-conditioned attention that injects camera-layout priors into graph self-attention for cross-view alignment.
• Leverage a Temporal Graph Network (TGN) for temporal and cross-camera motion modeling.
• Apply differentially private embedding release with calibrated Gaussian noise and use compact approximate indexes for secure retrieval.
• Integrate transport-regularized retrieval to encourage global, database-aware matching.

Figure 1: Overview of the CityGuard framework for bias-resilient, privacy-preserving identity search. The process begins with Topology-Aware Geometry Encoding , where camera coordinates and rotations are mapped to a spatial adjacency graph. The Geometry-Conditioned Backbone then fuses multi-scale fe

実験結果

リサーチクエスチョン

• RQ1 How can we achieve bias-resilient Re-ID across heterogeneous camera views under privacy constraints?
• RQ2 Can geometry-derived priors be effectively incorporated into graph-based attention to improve cross-view alignment without precise calibration?
• RQ3 How do adaptive margins based on per-identity dispersion influence intra-class compactness and cross-camera discrimination?
• RQ4 What is the impact of differentially private embedding releases on retrieval accuracy and scalability in urban surveillance settings?

主な発見

• CityGuard yields state-of-the-art or near state-of-the-art retrieval performance on standard benchmarks (e.g., Market-1501, MARS) when incorporating ACT loss, geometry-conditioned attention, and OT regularization.
• Geometry-aware attention improves cross-view alignment and robustness to viewpoint/occlusion beyond a geometry-free baseline.
• Adaptive dispersion-aware margins (ACT) enhance intra-class compactness and cross-camera discrimination, especially when combined with spatial priors.
• Differentially private embedding releases enable secure, scalable indexing with controllable privacy budgets, maintaining competitive retrieval performance under DP.
• Adversarial robustness tests show CityGuard maintaining higher Rank-1/mAP under FGSM and PGD-20 attacks compared with several baselines.
• Ablation studies indicate each CityGuard component (ACT, geo-attention, OT) contributes to performance gains, with the full framework delivering the strongest results.

Figure 2: Camera topology (GPS only): top-down 2D layout of camera nodes with edge thickness encoding the row-stochastic affinity $A_{ij}$ . This panel visualizes affinity derived solely from pairwise GPS distances.