Results of the summarizing process for the arXiv paper: 2306.06123v1

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.

• - 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

SummaryExplainable AI (XAI) helps us understand and trust computer models. But some bad people can trick the explanations. Researchers are studying how to make XAI more secure and fair. They want to protect against attacks that could change the explanations in a bad way. Definitions- Explainable AI (XAI): Methods that help us understand how computer models work. - Adversarial machine learning: Techniques used to trick or manipulate machine learning models. - Vulnerabilities: Weaknesses or flaws in something that can be exploited. - Manipulation: Changing something in a dishonest or unfair way. - Resilient: Able to withstand or recover from difficult situations.

In recent years, the field of artificial intelligence (AI) has seen a surge in interest and development. With advancements in statistical and deep learning models, AI has become increasingly capable of making complex decisions and predictions. However, as these models become more sophisticated, they also become less transparent to human understanding. This lack of transparency can lead to mistrust and skepticism towards AI systems, especially when they are used in critical decision-making processes. To address this issue, researchers have turned to explainable AI (XAI) methods as a solution for debugging and fostering trust in machine learning models. These methods aim to provide insights into how a model makes its predictions, allowing humans to understand the reasoning behind its decisions. However, recent advancements in adversarial machine learning have revealed limitations and vulnerabilities present in state-of-the-art explanations generated by XAI techniques. This raises concerns about the security and reliability of XAI systems. The potential for manipulation or deception of evidence regarding a model's reasoning poses significant risks when used in critical decision-making processes or knowledge discovery endeavors. To shed light on these issues, Hubert Baniecki and Przemyslaw Biecek have conducted a comprehensive survey that delves into the findings of over 50 research papers exploring adversarial attacks on explanations generated by machine learning models. Titled "Adversarial Attacks and Defenses in Explainable Artificial Intelligence: A Survey," this study 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 survey contributes valuable insights to ongoing discussions surrounding the robustness and integrity of interpretable AI technologies. The authors begin by providing an overview of explainable AI methods currently being used in various domains such as healthcare, finance, criminal justice system etc., highlighting their benefits but also acknowledging their limitations when it comes to handling adversarial attacks. They then delve into the various types of adversarial attacks that have been identified in the literature, including data poisoning, model inversion, and input perturbation attacks. These attacks aim to manipulate or deceive XAI systems by exploiting their vulnerabilities. The survey also considers fairness metrics within the context of adversarial attacks on XAI systems. This is an important aspect as these systems are often used in decision-making processes that can have significant impacts on individuals or groups. The authors discuss how adversarial attacks can lead to biased decisions and suggest ways to incorporate fairness considerations into defense mechanisms against such attacks. One of the key contributions of this study is its exploration of strategies for fortifying defenses against adversarial attacks on explanations generated by machine learning models. These include techniques such as robust feature selection, model distillation, and ensemble methods. The authors also discuss potential limitations and challenges associated with these defense mechanisms. In addition to discussing existing insecurities within XAI frameworks, the survey also sets the stage for outlining nascent avenues of inquiry within the domain of adversarial XAI (AdvXAI). It highlights areas where further research is needed to develop more robust and secure explainable AI methods. Overall, Baniecki and Biecek's survey sheds light on a critical issue facing interpretable AI technologies - their vulnerability to adversarial attacks. By providing a comprehensive overview of existing research in this area, it not only raises awareness about potential threats but also offers valuable insights into developing more resilient interpretation methodologies for safeguarding against malicious manipulations. In conclusion, while explainable AI methods have been hailed as a solution for fostering trust in machine learning models, they are not immune to security risks posed by adversarial attacks. This comprehensive survey serves as an important reminder that we must continue to critically examine and strengthen our understanding of AdvXAI if we want interpretable AI technologies to be reliable tools for decision-making processes in various domains.