Federated Active Learning for Target Domain Generalisation

Abstract: In this paper, we introduce Active Learning framework in Federated Learning for Target Domain Generalisation, harnessing the strength from both learning paradigms. Our framework, FEDALV, composed of Active Learning (AL) and Federated Domain Generalisation (FDG), enables generalisation of an image classification model trained from limited source domain client's data without sharing images to an unseen target domain. To this end, our FDG, FEDA, consists of two optimisation updates during training, one at the client and another at the server level. For the client, the introduced losses aim to reduce feature complexity and condition alignment, while in the server, the regularisation limits free energy biases between source and target obtained by the global model. The remaining component of FEDAL is AL with variable budgets, which queries the server to retrieve and sample the most informative local data for the targeted client. We performed multiple experiments on FDG w/ and w/o AL and compared with both conventional FDG baselines and Federated Active Learning baselines. Our extensive quantitative experiments demonstrate the superiority of our method in accuracy and efficiency compared to the multiple contemporary methods. FEDALV manages to obtain the performance of the full training target accuracy while sampling as little as 5% of the source client's data.

• The paper introduces Federated Domain Generalization (FDG), which minimizes feature complexity and condition alignment to tackle domain shifts.
• It integrates FEDALV, an active learning module that dynamically selects key client samples to maximize labeling efficiency.
• Experiments on PACS, OfficeHome, and OfficeCaltech demonstrate superior accuracy and efficiency while preserving data privacy.

Federated learning (FL) and active learning (AL) are two paradigms that have recently been making strides in improving machine learning models, particularly in scenarios where data privacy is crucial and labeled data is scarce. In “Federated Active Learning for Target Domain Generalization," the authors merge the two techniques (FL and AL) to address a specific challenge in the field: efficiently training models that are able to classify images even when they encounter data from new, previously unseen domains without having direct access to that data.

Traditional FL trains models on decentralized data, which is essential for user privacy. The models are updated locally on clients' devices, and only the model updates (not the raw data) are shared with a central server to maintain privacy. But one major drawback is the difficulty models have when facing 'domain shifts'—differences between the data they were trained on and new target data they encounter.

The paper introduces a novel approach named Federated Domain Generalization (FDG), which includes two key innovations for overcoming this challenge: loss functions designed to minimize 'feature complexity' and 'condition alignment' at the client level, and a 'regularization' mechanism at the server level that reduces biases between source (clients' data) and target domains (new, unseen data).

Moreover, the study proposes an AL component named FEDALV within the FDG framework, which actively selects the most informative samples from client data to label. As labeling data can be resource-intensive, finding and focusing on the most informative data points is critical. The FEDALV’s approach uses 'variable budgets,' allowing for flexibility in how many data points are queried from each client, enhancing the model’s ability to generalize by focusing on the most beneficial updates.

The authors conducted various experiments to compare their FDG and FEDALV frameworks against contemporary methods on image classification tasks using benchmark datasets like PACS, OfficeHome, and OfficeCaltech. The results demonstrated that their methods not only achieve superior accuracy but are also more efficient compared to other recent approaches.

The core of these advancements lies not just in effective model training but also in maintaining data privacy since the methods do not require sharing sensitive raw data. They exemplify the growing trend in AI research of finding innovative solutions that respect user privacy while improving model performance, an aspect increasingly relevant in today’s digitized world.

The authors also published their code publicly, which can be found on GitHub, promoting transparency, repeatability, and further innovation in the field. They stress the importance of future research including extending to a varied number of clients and domain heterogeneity, as well as the possibility of dynamic selection functions that further refine their approach.

In conclusion, the paper "Federated Active Learning for Target Domain Generalisation" marks a significant step forward in creating privacy-aware machine learning systems that are adaptable and accurate across a range of scenarios, reducing the need for direct data access and labor-intensive data labeling processes.