Certified Robustness for Deep Equilibrium Models via Serialized Random Smoothing

Authors: Weizhi Gao, Zhichao Hou, Han Xu, Xiaorui Liu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments and ablation studies on image recognition demonstrate that our algorithm can significantly accelerate the certification of DEQs by up to 7x almost without sacrificing the certified accuracy.
Researcher Affiliation Academia Weizhi Gao1 wgao23@ncsu.edu Zhichao Hou1 zhou4@ncsu.edu Han Xu2 xuhan2@arizona.edu Xiaorui Liu1* xliu96@ncsu.edu 1North Carolina State University, 2The University of Arizona *corresponding author
Pseudocode Yes Algorithm 1 Certified Radius with SRS-DEQ
Open Source Code Yes Our code is available at https://github.com/Weizhi Gao/Serialized-Randomized-Smoothing.
Open Datasets Yes Datasets. We use two classical datasets in image recognition, CIFAR-10 (Krizhevsky et al., 2009) and Image Net (Russakovsky et al., 2015), to evaluate the certified robustness.
Dataset Splits No The paper mentions training data and test data. It does not explicitly specify a validation split size or methodology. It discusses results based on different noise levels and number of layers, but not a validation set specifically.
Hardware Specification Yes All the experiments are conducted on one A100 GPU.
Software Dependencies No The paper does not list specific version numbers for software dependencies such as Python, PyTorch, or CUDA.
Experiment Setup Yes For the standard MDEQ on CIFAR10, we use the Anderson solver with the step of {1, 5, 30}. For the standard MDEQ on Image Net, we use the Broyden solver with the step of {1, 5, 14}. We apply Anderson and Naive solvers on CIFAR-10 and Broyden solver on Image Net for the proposed SRS-MDEQ with the step of {1, 3}. We adopt a warm-up technique, where we use multi-steps to solve the fixed-point problem for the first batch in Algorithm 1. The warm-up steps for our SRS-MDEQ are set as 30 and 14 steps for CIFAR-10 and Image Net, respectively. The failure rate as α = 0.001 and the sampling number as N = 10,000 in the Monte Carlo method, unless specified otherwise.