Incremental Randomized Smoothing Certification

Authors: Shubham Ugare, Tarun Suresh, Debangshu Banerjee, Gagandeep Singh, Sasa Misailovic

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally demonstrate the effectiveness of our approach, showing up to 4.1x certification speedup over the certification that applies randomized smoothing of approximate model from scratch. We extensively evaluate the performance of IRS when applying several common DNN approximations such as pruning and quantization on state-of-the-art DNNs on CIFAR10 (Res Net-20, Res Net-110) and Image Net (Res Net-50) datasets.
Researcher Affiliation Collaboration 1University of Illinois Urbana-Champaign, 2VMware Research
Pseudocode Yes Algorithm 1 RS certification (Cohen et al., 2019) Inputs: f: DNN, σ: standard deviation, x: input to the DNN, n0: number of samples to predict the top class, n: number of samples for computing p A, α: confidence parameter... Algorithm 2 IRS algorithm: Certification with cache Inputs: f p: DNN obtained from approximating f, σ: standard deviation, x: input to the DNN, np: number of Gaussian samples used for certification, Cf: stores the information to be reused from certification of f, α and αζ: confidence parameters, γ: threshold hyperparameter to switch between estimation methods... Algorithm 3 Estimate ζx Inputs: f p: DNN obtained from approximating f, σ: standard deviation, x: input to the DNN, np: number of Gaussian samples used for estimating ζx, Cf: stores the information to be reused from certification of f, αζ: confidence parameter Output: Estimated value of ζx
Open Source Code Yes IRS code is available at https://github.com/uiuc-arc/Incremental-DNN-Verification.
Open Datasets Yes We evaluate IRS on CIFAR-10 (Krizhevsky et al.) and Image Net (Deng et al., 2009).
Dataset Splits No The paper states using CIFAR-10 and Image Net datasets and refers to a 'test set' and 'validation set images' but does not provide explicit details on the training, validation, or test splits (e.g., percentages, sample counts, or specific predefined splits).
Hardware Specification Yes We ran experiments on a 48-core Intel Xeon Silver 4214R CPU with 2 NVidia RTX A5000 GPUs.
Software Dependencies Yes IRS is implemented in Python and uses Py Torch 2.0.1. (Paszke et al., 2019).
Experiment Setup Yes Hyperparameters. We use confidence parameters α = 0.001 for the certification of g, and αζ = 0.001 for the estimation of ζx. To establish a fair comparison, we set the baseline confidence with αb = α + αζ = 0.002. This choice ensures that both the baseline and IRS, provide certified radii with equal confidence. We use grid search to choose an effective value for γ. A detailed description of our hyperparameter search and its results are described in Section 5.4. ... We use n = 105 samples for certification of g. For certifying gp, we consider np values from {5%, . . . 50%} of n and σ = 1.