Enhanced Regularizers for Attributional Robustness

Authors: Anindya Sarkar, Anirban Sarkar, Vineeth N Balasubramanian2532-2540

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted a comprehensive suite of experiments and ablation studies, which we report in this section and in Sec . We report results with our method on 4 benchmark datasets i.e. Flower (Nilsback and Zisserman 2006), Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), MNIST and GTSRB (Stallkamp et al. 2012).
Researcher Affiliation Collaboration Anindya Sarkar, Anirban Sarkar, Vineeth N Balasubramanian Indian Institute of Technology, Hyderabad anindya.sarkar@cse.iith.ac.in, cs16resch11006@iith.ac.in, vineethnb@ith.ac.in
Pseudocode Yes An algorithm for our overall methodology is also presented in the Appendix due to space constraints.
Open Source Code No The paper does not contain an explicit statement about releasing source code for the described methodology or a direct link to a code repository.
Open Datasets Yes We report results with our method on 4 benchmark datasets i.e. Flower (Nilsback and Zisserman 2006), Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), MNIST and GTSRB (Stallkamp et al. 2012).
Dataset Splits No The paper mentions training data and test sets but does not explicitly provide details on validation splits (e.g., percentages, sample counts, or clear references to predefined validation sets).
Hardware Specification No The paper mentions 'GPU servers' for provision of work but does not provide specific details such as exact GPU models, CPU models, or memory specifications used for running experiments.
Software Dependencies No The paper describes the network architectures used but does not provide specific software dependency details with version numbers (e.g., Python, PyTorch/TensorFlow versions, or other libraries).
Experiment Setup Yes We used a regularizer coefficient λ = 1.0 and m = 50 as the number of steps used for computing IG (Eqn 1) across all experiments. Note that our adversarial and attributional attack configurations were kept fixed across ours and baseline methods. Please refer the Appendix for more details on training hyperparameters and attack configurations for specific datasets.