This is the introduction and repository for XAIFooler the algorithm behind "Are Your Explanations Reliable?" Investigating the Stability of LIME in Explaining Text Classifiers by Marrying XAI and Adversarial Attack" (EMNLP 2023).
XAIFooler is built on the TextAttack Framework and uses the ELI5 implementation of LIME. It is recommended to install both packages in a dedicated environment through Anaconda or similar. Installations on Windows systems may encounter issues with automatic installation of ELI5 due to deprecated packages. Linux installations have not been problematic.
Two options are available, the most user friendly is the included python notebook. Easier access to altering the parameters (see below) can be done through using the main.py file to pass them directly as arguments on the command line. Adjustments should be made to the batch size depending of available computational resources. Currently overrided to work on low resource machines.
The possible arguments for adjusting the adversarial search process can be seen directly in common.py and are as follows:
- Method - Which search method to implement, options are XAIFooler, Inherent (None), Random (Greedy), Random (Non-Greedy), GA (Genetic Algorithm)
- LIME's sampling rate (ELI5 has the default implementation at 5000, see the paper for more efficient values for particular datasets, for the datasets used in the paper they are set to the values seen there by default)
- The top-n feature threshold - Which features are required to be held constant between each iteration of the search
- Max Candidates - The number of nearest neighbors calculated for word replacement
- Modify Rate - The maximum percentage of possible perturbed words
- num - The number of documents to attack (Reccomended due to time only to choose a small subset of the included datasets)
- Success Threshold - Not used in the paper's experiments, this sets a similarity threshold to allow early termination of the search once a desired dissimilarity between the original explanation and the perturbed explanation is reached.
- Similarity Measure - Which measure to use for explanation comparison during the search process (Options are RBO, L2, Center of Mass (COM), Center of Mass (Rank Weighted), Jaccard, Kendall, Spearman, Jaccard (Weighted), Spearman (Weighted), Kendall (Weighted)
- RBO-p - for the RBO similarity measure, the weight parameter that controls the top-weightedness
- Random Seed
- Model - BERT, RoBERTa, DistilBERT versions fine-tuned on the respective dataset
- Dataset - IMDB, Symptoms to Diagnosis, Twitter Hate Speech
- Genetic Algorithm Specific Flags - Crossover and Parent Selection
@inproceedings{ xaifooler, title={``Are Your Explanations Reliable?" Investigating the Stability of LIME in Explaining Text Classifiers by Marrying XAI and Adversarial Attack}, author={Christopher Burger, Lingwei Chen, Thai Le}, booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing}, year={2023} }