BagFlip: A Certified Defense Against Data Poisoning

Authors: Yuhao Zhang, Aws Albarghouthi, Loris D'Antoni

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Bag Flip on image classification and malware detection datasets. Bag Flip is equal to or more effective than the state-of-the-art approaches for trigger-less attacks and more effective than the state-of-the-art approaches for backdoor attacks.
Researcher Affiliation Academia Yuhao Zhang University of Wisconsin-Madison yuhaoz@cs.wisc.edu Aws Albarghouthi University of Wisconsin-Madison aws@cs.wisc.edu Loris D Antoni University of Wisconsin-Madison loris@cs.wisc.edu
Pseudocode No The paper describes its algorithms and mathematical formulations in prose and equations, but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The implementation of Bag Flip is publicly available2. https://github.com/Forever Zyh/defend_framework
Open Datasets Yes We conduct experiments on MNIST, CIFAR10, EMBER [2], and Contagio (http://co ntagiodump.blogspot.com).
Dataset Splits No The paper mentions training and testing but does not explicitly describe validation dataset splits, percentages, or methodology for a validation set.
Hardware Specification No The paper mentions 'a single GPU' and 'a single core' for training and preparation, but it does not specify the model or type of these hardware components (e.g., NVIDIA A100, Intel Xeon).
Software Dependencies No The paper mentions training 'neural networks' and 'random forests' but does not specify any software libraries or frameworks (e.g., PyTorch, TensorFlow, Scikit-learn) along with their version numbers.
Experiment Setup Yes We train N = 1000 models and set the confidence level as 0.999 for each configuration. ... We set k = 100, 1000, 300, 30 for MNIST, CIFAR10, EMBER, and Contagio respectively when comparing to Bagging. We tune k = 80, 280 for Bagging-0.9 on MNIST and Bagging-0.95 on EMBER, respectively. And we set k = 50 for MNIST when comparing to Label Flip.