Certified Neural Network Watermarks with Randomized Smoothing

Authors: Arpit Bansal, Ping-Yeh Chiang, Michael J Curry, Rajiv Jain, Curtis Wigington, Varun Manjunatha, John P Dickerson, Tom Goldstein

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

Reproducibility Variable Result LLM Response
Research Type Experimental In our first set of experiments, we investigate the strength of our certificate under two datasets and three watermark schemes. In our second set of experiments, we evaluate the watermark s empirical robustness to removal compared to previous methods that claimed resistance to removal attacks.
Researcher Affiliation Collaboration 1University of Maryland, College Park 2Adobe Research, USA.
Pseudocode Yes Algorithm 1 Embed Certifiable Watermark Algorithm 2 Evaluate and Certify the Median Smoothed Model
Open Source Code No The paper does not contain an explicit statement about releasing the source code for their method, nor does it provide a link to a code repository.
Open Datasets Yes For example, MNIST images could form the trigger set for a CIFAR-10 network. To train the watermarked model, we used Res Net-18. We used the Watermark-Robustness-Toolbox to conduct the additional persistence evaluation.
Dataset Splits No No explicit train/validation/test splits were provided beyond 'Only 50% of the data is used for training, since we reserve the other half for the adversary.'
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU or CPU models. It mentions 'TPU cost' in reference to GPT-3 but not for their own experimental setup.
Software Dependencies No The paper does not provide specific software dependency details with version numbers. It mentions using 'ResNet-18' and optimizers like 'SGD' and 'Adam', but not the software framework or library versions used for implementation.
Experiment Setup Yes To train the watermarked model, we used Res Net-18, SGD with learning rate of .05, momentum of .9, and weight decay of 1e-4. The model is trained for 100 epochs, and the learning rate is divided by 10 every 30 epochs. ... For our watermark models, we select σ of 1, replay count of 20, and noise sample count of 100. ... To attack the model, we used Adam with learning rates of .1, .001 or .0001 for 50 epochs.