Title: AR2PNET: an adversarially robust re-weighting prototypical network for few-shot learning

Authors: Sirui Li; Li Guo; Xianmin Wang; Songcao Hou; Zhicong Qiu; Yutong Xie; Haiyan Liang

Addresses: Institute of Artificial Intelligence and Blockchain, Guangzhou, Guangdong, China ' Zhongshan Road Primary School, Ganzhou, Jiangxi, China ' Institute of Artificial Intelligence and Blockchain, Guangzhou University, Guangzhou, Guangdong, China ' Institute of Artificial Intelligence and Blockchain, Guangzhou University, Guangzhou, Guangdong, China ' Institute of Artificial Intelligence and Blockchain, Guangzhou University, Guangzhou, Guangdong, China ' Institute of Artificial Intelligence and Blockchain, Guangzhou University, Guangzhou, Guangdong, China ' Institute of Artificial Intelligence and Blockchain, Guangzhou University, Guangzhou, Guangdong, China

Abstract: Robust re-weighting prototypical networks (RRPNet) model is a promising method to improve the robustness of prototypical networks (ProtoNet). However, the performance of RRPNet is limited when the examples are scare and the noise is trivial. In this paper we propose a novel re-weighting prototypical networks framework for few-shot learning based on AT, called AR2PNet, to enhance the performance of RRPNet. Specifically, instead of directly calculating the similarity between the naive representations of the examples, we calculate such similarity between prototype representations, which is conductive to reducing the computation cost as well as enhancing the model prediction accuracy. Meanwhile, to encourage the model to resist adversarial examples, we formulate the loss function as a minimax problem inspired by the conception of AT. We conduct experiments on CIFAR-FS and MiniImageNet dataset, and the experimental results demonstrate the effectiveness of the propose method.

Keywords: deep learning; few-shot learning; prototypical network; adversarial training; AT.

DOI: 10.1504/IJES.2023.136376

International Journal of Embedded Systems, 2023 Vol.16 No.2, pp.96 - 104

Received: 01 Dec 2022
Accepted: 21 Mar 2023

Published online: 31 Jan 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article