Adversarial Attacks on Gaussian Process Bandits
Authors: Eric Han, Jonathan Scarlett
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that adversarial attacks on GP bandits can succeed in forcing the algorithm towards Rtarget even with a low attack budget, and we test our attacks effectiveness on a diverse range of objective functions. ... We demonstrate the effectiveness of these attacks via experiments on a diverse range of objective functions. |
| Researcher Affiliation | Academia | School of Computing, National University of Singapore Department of Mathematics & Institute of Data Science, National University of Singapore. |
| Pseudocode | No | The paper describes algorithms and attack methods using prose and mathematical equations (e.g., equations 5, 12, 13, 14) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/eric-vader/Attack-BO. |
| Open Datasets | Yes | Implementations of the synthetic functions are from HPOlib2 (Eggensperger et al., 2013). We use the original authors implementation (Wang & Jegelka, 2017) for the robot pushing objective functions (Robot3D and Robot4D). |
| Dataset Splits | No | The paper describes the number of initial points and iterations for the bandit optimization but does not provide traditional training/validation/test dataset splits as data is sampled sequentially by the bandit algorithm. |
| Hardware Specification | No | Our implementation is based on Python 3.8, using Conda (Anaconda, 2016) and MLflow (Zaharia et al., 2018) to manage environments and experiments across multiple machines." This statement does not specify any particular hardware. |
| Software Dependencies | Yes | Our implementation is based on Python 3.8, using Conda (Anaconda, 2016) and MLflow (Zaharia et al., 2018) to manage environments and experiments across multiple machines. Our implementation uses several standard machine learning and scientific packages, such as GPy (GPy, 2012), Num Py (Harris et al., 2020) and others. The various libraries and the specific versions used can be found in our supplied code, in the Conda environment file attack bo.yml. |
| Experiment Setup | Yes | We run each attack 300 times, varying 30 different hyperparameter choices (i.e. , hmax, pµa, σaq, hpxq) with 10 differing conditions (initial points, instances of the objective function, and random seeds). ...We run 100 or 250 optimization steps depending on the dimensionality... We select the exploration parameter as βt 0.5 log p2tq. |