A Zest of LIME: Towards Architecture-Independent Model Distances

Authors: Hengrui Jia, Hongyu Chen, Jonas Guan, Ali Shahin Shamsabadi, Nicolas Papernot

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that our method, which we call Zest, helps in several tasks that require measurements of model similarity: verifying machine unlearning, and detecting many forms of model reuse, such as model stealing, knowledge distillation, and transfer learning.
Researcher Affiliation Collaboration Hengrui Jia, Hongyu Chen, Jonas Guan University of Toronto and Vector Institute {nickhengrui.jia, hy.chen}@mail.utoronto.ca, jonas@cs.toronto.edu Ali Shahin Shamsabadi Vector Institute and The Alan Turing Institute a.shahinshamsabadi@turing.ac.uk Nicolas Papernot University of Toronto and Vector Institute nicolas.papernot@utoronto.ca
Pseudocode Yes Algorithm 1: Proposed model distance approach, Zest
Open Source Code Yes Code is available at: https://github.com/cleverhans-lab/Zest-Model-Distance
Open Datasets Yes We validate our proposed method, Zest, in the vision, text, and audio domains using four publicly available classifiers and four datasets: Res Net20 trained on CIFAR-10, Res Net50 trained on CIFAR-100, LSTM trained on AG News, and M5 network trained on Speech Commands.
Dataset Splits No The paper uses standard datasets like CIFAR-10 and CIFAR-100, which have predefined splits, but does not explicitly state the train/validation/test splits (e.g., percentages or exact counts) within the paper's text for the models themselves.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using open-source implementations for models and algorithms (e.g., Res Net, LIME), implying software like PyTorch, but does not provide specific version numbers for these software components.
Experiment Setup Yes In phase 1 of Zest, we select 128 samples randomly from the training set. In phase 1 and phase 2, we follow the same implementation and setup as the original LIME paper. In phase 3, we train linear regression models using the least-squares estimation technique. We use relative Cosine similarity (see Appendix C for details) for computing the distance between global signatures in phase 4. Following the original LIME paper, we set L = 1000.