# Adversarial Attacks and Defenses in Explainable Artificial Intelligence: A Survey

**Authors:** Hubert Baniecki, Przemyslaw Biecek  
**arXiv:** [2306.06123](https://arxiv.org/abs/2306.06123)  
**Date:** 2023-06-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.06123v1/

## Key points

- Explainable AI (XAI) methods are seen as a solution for debugging and building trust in statistical and deep learning models
- XAI methods offer insights into model predictions but recent advancements in adversarial machine learning have exposed limitations and vulnerabilities in explanations
- Concerns arise about the security and reliability of XAI explanations due to potential manipulation, deception, or whitewashing of evidence
- A comprehensive survey based on over 50 research papers explores adversarial attacks on machine learning model explanations and considers fairness metrics
- Strategies are discussed for strengthening defenses against attacks and developing resilient interpretation methodologies to prevent malicious manipulations
- The survey highlights existing insecurities within XAI frameworks, paving the way for further exploration in adversarial XAI (AdvXAI)
- Authored by Hubert Baniecki and Przemyslaw Biecek, the survey titled "Adversarial Attacks and Defenses in Explainable Artificial Intelligence: A Survey" will be presented at the IJCAI 2023 Workshop on XAI

## Summary

In the realm of artificial intelligence, explainable AI (XAI) methods have been heralded as a solution for debugging and fostering trust in statistical and deep learning models. These methods also provide insights into their predictions. However, recent advancements in adversarial machine learning have revealed limitations and vulnerabilities present in state-of-the-art explanations. This casts doubt on their security and reliability. The potential for manipulation, deception, or whitewashing of evidence regarding a model's reasoning poses significant risks when used in critical decision-making processes and knowledge discovery endeavors. This comprehensive survey delves into the findings of over 50 research papers that explore adversarial attacks on explanations generated by machine learning models. It also considers fairness metrics within these contexts. The discourse extends to strategies for fortifying defenses against such attacks and devising resilient interpretation methodologies to safeguard against malicious manipulations. By identifying a spectrum of existing insecurities within XAI frameworks, the survey sets the stage for outlining nascent avenues of inquiry within the domain of adversarial XAI (AdvXAI). Authored by Hubert Baniecki and Przemyslaw Biecek, this survey titled "Adversarial Attacks and Defenses in Explainable Artificial Intelligence: A Survey" is slated for presentation at the IJCAI 2023 Workshop on XAI. Through its nuanced examination of adversarial threats to XAI systems and proposed defense mechanisms, this study contributes valuable insights to ongoing discussions surrounding the robustness and integrity of interpretable AI technologies.