Federated Learning of Low-Rank One-Shot Image Detection Models in Edge Devices with Scalable Accuracy and Compute Complexity

TL;DR

LoRa-FL tackles efficient federated training of one-shot image detection on resource-limited edge devices by updating only low-rank adapters in a Siamese detector, while freezing backbone weights. It implements $W_i = A_i B_i$ with rank $k$ to minimize communication and compute without sacrificing detection performance. Experiments on MNIST and CIFAR-10 under IID and non-IID conditions show favorable accuracy with moderate $k$ and substantial resource savings, demonstrating scalable deployment for heterogeneous edge environments. The work provides a practical, open-source framework for privacy-preserving, low-rank federated learning in compact vision tasks.

Abstract

This paper introduces a novel federated learning framework termed LoRa-FL designed for training low-rank one-shot image detection models deployed on edge devices. By incorporating low-rank adaptation techniques into one-shot detection architectures, our method significantly reduces both computational and communication overhead while maintaining scalable accuracy. The proposed framework leverages federated learning to collaboratively train lightweight image recognition models, enabling rapid adaptation and efficient deployment across heterogeneous, resource-constrained devices. Experimental evaluations on the MNIST and CIFAR10 benchmark datasets, both in an independent-and-identically-distributed (IID) and non-IID setting, demonstrate that our approach achieves competitive detection performance while significantly reducing communication bandwidth and compute complexity. This makes it a promising solution for adaptively reducing the communication and compute power overheads, while not sacrificing model accuracy.

Figures (7)

• Figure 1: Scenario considered in this work. A number of edge devices with limited compute capacity are considered within a Federated Learning (FL) setting, where the goal is to learn a one-shot image detection model in a federated manner. Given a target image, the one-shot model can predict if an incoming query image is of the same class as the target or not. Within this context, we propose an approach termed LoRa-FL for adaptively reducing the edge-server communication bandwidth requirements as well as the compute complexity needed for the edge devices during the FL process. Our approach makes use of Low-Rank (LoRa) modules within the one-shot network architecture which effectively reduces the number of trainable weights by modulating the rank $k$ of the LoRa modules.
• Figure 2: One-shot Siamese model architecture considered in this work. The target and query images are fed to their respective pre-trained SqueezeNet backbone, acting as a Siamese feature extraction module. The target and query feature vectors $x_1$ and $x_2$ are then processed by two cascaded LoRa layers of rank $k$ (where $k$ varies during our experiments). The produced target and query embeddings $u_1$ and $u_2$ are then used to compute the absolute difference vector $|u_1 - u_2|$ which is in turn processed by a final LoRa module. Finally, the output of the similarity_LoRa module is processed by a Sigmoid layer with one output indicating similarity or dissimilarity between the target and query images.
• Figure 3: MNIST IID: Test accuracy in function of the global FL round.
• Figure 4: CIFAR10 IID: Test accuracy in function of the global FL round.
• Figure 5: MNIST NON-IID: Test accuracy in function of the global FL round.