Privacy-Preserving Adaptive Re-Identification without Image Transfer

TL;DR

Privacy constraints in public deployments hinder centralized Re-ID training. The authors propose Fed-Protoid, a prototype-based federated approach for DUDA-Rid that uses source prototypes and a distributed $MMD$ loss to align edge-device features without sharing images. A teacher-student framework plus a pseudo-client trained on the source domain enables cross-domain adaptation, and Fed-Protoid++ with ViT and self-supervised pretraining on LUPerson yields further gains. Experiments on real-to-real and synthetic-to-real benchmarks show Fed-Protoid outperforms adapted UDA and federated baselines while reducing communication. The work demonstrates a practical, privacy-preserving pathway for adaptive Re-ID in real-world deployments.

Abstract

Re-Identification systems (Re-ID) are crucial for public safety but face the challenge of having to adapt to environments that differ from their training distribution. Furthermore, rigorous privacy protocols in public places are being enforced as apprehensions regarding individual freedom rise, adding layers of complexity to the deployment of accurate Re-ID systems in new environments. For example, in the European Union, the principles of ``Data Minimization'' and ``Purpose Limitation'' restrict the retention and processing of images to what is strictly necessary. These regulations pose a challenge to the conventional Re-ID training schemes that rely on centralizing data on servers. In this work, we present a novel setting for privacy-preserving Distributed Unsupervised Domain Adaptation for person Re-ID (DUDA-Rid) to address the problem of domain shift without requiring any image transfer outside the camera devices. To address this setting, we introduce Fed-Protoid, a novel solution that adapts person Re-ID models directly within the edge devices. Our proposed solution employs prototypes derived from the source domain to align feature statistics within edge devices. Those source prototypes are distributed across the edge devices to minimize a distributed Maximum Mean Discrepancy (MMD) loss tailored for the DUDA-Rid setting. Our experiments provide compelling evidence that Fed-Protoid outperforms all evaluated methods in terms of both accuracy and communication efficiency, all while maintaining data privacy.

Figures (6)

• Figure 1: In traditional Unsupervised Domain Adaptation (UDA) as depicted in Fig. (a), images are transmitted to a centralized server, which combines the unlabeled target images with the annotated source samples to train a model. In contrast, Distributed UDA for person re-identification (DUDA-Rid) shown in Fig. (b) keeps target images exclusively on edge devices. The learning process is divided between the server and cameras, the latter being equipped with local computational resources (). Only model parameters are exchanged between the clients and the server.
• Figure 1: The impact of the number of source prototypes in the Fed-Protoid performance in two configurations: RP$\to$M and RP$\to$C
• Figure 2: The pipeline of Fed-Protoid. Our algorithm aggregates $n$ edge-client models and one pseudo-client model in the server. Source prototypes are computed with the aggregated model. The prototypes and aggregated model are then distributed to all edge devices for local unsupervised training. This local training on each client involves cross-entropy, triplet, and Maximum Mean Discrepancy (MMD) loss functions.
• Figure 2: Impact of the backbone architecture and pre-training datasets on Fed-Protoid.
• Figure 2: Ablation study on the sensibility of the different hyper-parameters of Fed-Protoid.