Training for Stable Explanation for Free

Authors: Chao Chen, Chenghua Guo, Rufeng Chen, Guixiang Ma, Ming Zeng, Xiangwen Liao, Xi Zhang, Sihong Xie

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across various data modalities and model architectures show that R2ET achieves superior stability against stealthy attacks, and generalizes effectively across different explanation methods.
Researcher Affiliation Collaboration 1School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), China 2Key Laboratory of Trustworthy Distributed Computing and Service (Mo E), Beijing University of Posts and Telecommunications, China 3Artificial Intelligence Thrust, The Hong Kong University of Science and Technology (Guangzhou), China 4 Intel, USA 5Carnegie Mellon University, USA 6College of Computer and Data Science, Fuzhou University, China
Pseudocode Yes Algorithm 1 Attacking a pair of features
Open Source Code Yes The code can be found at https://github.com/ccha005/R2ET.
Open Datasets Yes We adopt DNNs for three tabular datasets: Bank [53], Adult, and COMPAS [54], and two image datasets, CIFAR-10 [36] and ROCT [22]... dual-input Siamese networks [48] are adopted on MNIST [40] and two graph datasets, BP [48] and ADHD [49].
Dataset Splits Yes We divide each tabular dataset into training, validation, and test portions at a ratio of 70 : 15 : 15, respectively.
Hardware Specification Yes Both come with a 16-core Intel Xeon processor and four TITAN X GPUs. One installs 16.04.3 Ubuntu with 3.8.8 Python and 1.7.1 Py Torch, and the other installs 18.04.6 Ubuntu with 3.7.6 Python and 1.8.1 Py Torch. The MNIST and graph datasets are run on a machine with two 10-core Intel Xeon processors and five Ge Force RTX 2080 Ti GPUs, which installs 18.04.3 Ubuntu with 3.9.5 Python and 1.9.1 Py Torch.
Software Dependencies Yes One installs 16.04.3 Ubuntu with 3.8.8 Python and 1.7.1 Py Torch, and the other installs 18.04.6 Ubuntu with 3.7.6 Python and 1.8.1 Py Torch. The MNIST and graph datasets are run on a machine with two 10-core Intel Xeon processors and five Ge Force RTX 2080 Ti GPUs, which installs 18.04.3 Ubuntu with 3.9.5 Python and 1.9.1 Py Torch.
Experiment Setup Yes Table 4: (Hyper)-parameters used in experiments. SN means Siamese Networks for dual input.