RetrievalGuard: Provably Robust 1-Nearest Neighbor Image Retrieval

Authors: Yihan Wu, Hongyang Zhang, Heng Huang

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on image retrieval tasks validate the robustness of our Retrieval Guard method.
Researcher Affiliation Academia 1Department of Electrical and Computer Engineering, University of Pittsburgh, USA 2David R. Cheriton School of Computer Science, University of Waterloo, Canada.
Pseudocode Yes Algorithm 1 Certified Radius by Retrieval Guard
Open Source Code No The paper does not provide any links to source code or explicitly state that code is made publicly available.
Open Datasets Yes We run experiments with a popular dataset Online Products of metric learning (Song et al., 2016)... The experiments with CUB200 (Wah et al., 2011) and CARS196 (Krause et al., 2013) are listed in Appendix C.
Dataset Splits No The paper specifies training and test set splits for the datasets but does not explicitly mention or provide details for a validation set split.
Hardware Specification Yes The running time of Retrieval Guard for a single image evaluation with 100,000 Monte-Carlo samples on a 24GB Nvidia Tesla P40 GPU is about 3 minutes.
Software Dependencies No The paper mentions using "ResNet50 architecture" and adapting a "DML framework from (Roth et al., 2020)" but does not specify version numbers for any software dependencies like PyTorch, TensorFlow, Python, or CUDA.
Experiment Setup Yes The embedding dimension k is 128 and the number of training epochs is 100. The learning rate is 1e-5 with multi-step learning rate scheduler 0.3 at the 30-th, 55-th, and 75-th epochs. We select the initial value of β as 1.2, learning rate of β as 0.0005, and γ = 0.2 in the margin loss. We also test the model performance under Gaussian noise with σ = 0.1, 0.25, 0.5, 1. As the embedding of metric learning models is ℓ2 normalized, we have F = 1. For each sample, we generate 100,000 Monte-Carlo samples to estimate g(x). The confidence level α is chosen as 0.01.