中文

ANROT-HELANet:基于希洛inger距离的对抗鲁棒注意力聚合网络

计算机视觉与模式识别 2025-09-16 v1

摘要

few-shot 学习(FSL)涉及仅用少量数据样本即可实现的泛化学习,已显示出优于普通卷积神经网络方法的优异性能。虽然基于贝叶斯的方法使用Kullback-Leibler(KL)散度已取得改进,但仍然容易受到对抗攻击和自然噪声的威胁。我们引入ANROT-HELANet,这是一个 Adversarially and Naturally RObust Hellinger Aggregation Network,显著推动了FSL鲁棒性和性能的前沿。我们的方法实现了一种对抗鲁棒且自然鲁棒的希洛inger距离-based feature class aggregation scheme,显示出对对抗扰动(up to ϵ=0.30\epsilon=0.30)和高斯噪声(up to σ=0.30\sigma=0.30)的韧性。该网络在基准数据集上实现了显著的改进,包括在miniImageNet上分别获得1-shot和5-shot情景下的1.20%和1.40%的提升。我们提出了 novel Hellinger Similarity contrastive loss 函数,generalizes cosine similarity contrastive loss 用于变分few-shot 推理情景。我们的方法还在图像重建质量方面取得优异成绩,FID得分为2.75,优于传统VAE(3.43)和WAE(3.38)方法。广泛的实验在四个few-shot benchmark 数据集上进行,证明ANROT-HELANet 结合希洛inger距离-based feature aggregation、注意力机制和我们 novel loss 函数在保持对抗和自然扰动鲁棒性的同时,建立了新的 state-of-the-art performance。我们的代码仓库将在 https://github.com/GreedYLearner1146/ANROT-HELANet/tree/main 提供。

关键词

引用

@article{arxiv.2509.11220,
  title  = {ANROT-HELANet: Adverserially and Naturally Robust Attention-Based Aggregation Network via The Hellinger Distance for Few-Shot Classification},
  author = {Gao Yu Lee and Tanmoy Dam and Md Meftahul Ferdaus and Daniel Puiu Poenar and Vu N. Duong},
  journal= {arXiv preprint arXiv:2509.11220},
  year   = {2025}
}

备注

Preprint version. The manuscript has been submitted to a journal. All changes will be transferred to the final version if accepted. Also an erratum: In Figure 10 and 11, the $\epsilon = 0.005$ value should be $\epsilon = 0.05$