Transparency in Artificial Intelligence: An Overview

ABSTRACT

This research paper investigates the concept of transparency in artificial intelligence (AI). With the increase of AI in high-stake industries such as healthcare, finance, telecommunications, energy, and aviation, the importance of transparency in explainable artificial intelligence (XAI) becomes increasingly important to understand the decision-making processes of AI. This study will explore past and current techniques of AI and evaluate the effects of their respective interpretability and explainability capabilities. The research will also address the key considerations of bias, fairness, and trust about their impact on accuracy, reliability, and acceptability. The findings of this study will help further the understanding of the XAI’s role in ensuring the responsible and ethical use of AI.

1 INTRODUCTION

Transparency has emerged as a critical issue in the field of artificial intelligence (AI) as concerns around the interpretability, explainability, and accountability of these systems become more prevalent.

One of the main concerns around AI systems is that it can be difficult for humans to understand how decisions are made. This lack of transparency can have significant ethical implications when these systems are used in high-stakes decision-making contexts such as healthcare. We will provide a brief overview of the history of AI and how it has evolved, including the factors that contributed to the current black-box nature of many AI systems. Then, we will examine the use of AI in various industries including healthcare and companies like Amazon and Twitter. We will highlight the ethical risks and potential and actual consequences of a lack of transparency. We will also review existing literature and technologies revolving around AI transparency challenges and solutions including various model interpretation methods and tools that have been developed to address the interpretability and explainability of AI systems. Additionally, we will discuss the effects it has on privacy.

We hope to provide valuable insights for individual users, companies, and AI researchers about the importance of transparency in AI systems.

2 HISTORY

In the 1950s, the concept of AI was inspired by science fiction novels and the ideas put forth by researchers such as Turing. At that time, computers were expensive and could only execute commands without being able to store them. However, with advancements in technology, the beginnings of modern computers and AI flourished from 1957–1974 as computers became more accessible, faster, and cheaper. During this time, various definitions for AI emerged ranging from its ability to perform specific tasks through flexible adaptation to learning and generalizing like humans.

In the early days of AI, the focus was on developing systems that could perform specific tasks, such as playing games or recognizing patterns in data. The inner workings of these systems were not considered important, as long as they produced accurate results. Today, AI is often referred to as a “black-box” technology, where its internal workings and decision-making processes are opaque to human understanding. This lack of transparency is due to the complexity of the algorithms and the sheer volume of data used by these systems. This makes interpretation and understanding how they arrive at their decisions very difficult.

3 TRANSPARENCY

3.1 MOTIVATION

Whether we know it or not, AI is everywhere around us. “The global artificial intelligence market was valued at USD 136.55 billion in 2022.” AI has shown to be a significant revolution. $36.4B of the previously mentioned share was attributed to deep learning. [1] While deep learning is proven to be an extremely useful, cost-saving, and revenue-driving technology, it is considered a black box. Their complex interplay between neurons and weights makes it impossible to trace or interpret its decisions leading to an output. In industry, confidence in understanding and trust is extremely important. We need to be able to see why an AI is deciding because it could give a bad result based on the wrong reasons and unforeseen biases.

For the end-user, it is important to understand why an AI is making a particular decision or prediction. By providing an explainable and interpretable system, users can have more trust. This leads to increased adoption and use.

For industries, such as healthcare, finance, and tech, AI is important for ensuring decisions made by these algorithms are correct, fair, and ethical. The lack of transparency makes it difficult to detect biases, errors, or other issues that could lead to negative outcomes.

For researchers, developing transparent AI can lead to new insights into how these systems work and how to improve on them. This will lead to more trustworthy and responsible AI.

3.2 DEFINITION

There is no standard and generally accepted definition of explainable AI. However, to put some clarification around the concept, XAI aims to “produce more explainable models, while maintaining a high level of learning performance (prediction accuracy); and enable human users to understand, appropriately, trust, and effectively manage the emerging generation of artificially intelligent partners.” [2] We can break this down with a few key terms:

• Explainability: The ability of an AI system to provide humans with understandable explanations of its decision-making process and how it arrived at a particular output.
• Interpretability: The ability of an AI system to be understood and interpreted by humans. Experts in the field must be able to ensure that the system is working as intended.
• Accountability: The ability of relevant stakeholders, such as users, auditors, and regulators, to access and review the data and algorithms behind an AI system to ensure the system is providing trustworthy and ethical outputs.

3.3 BIASES and TRUST

Design guidelines for AI models have traditionally prioritized maximizing accuracy. However, recent research has demonstrated that neglecting issues of discrimination and unfairness may result in scenarios that are economically poor and socially unacceptable. There is growing recognition that these issues need to be explicitly addressed in designing AI models. AI should be accountable from an ethical view. [3] AI development and decision-making can be roughly represented as a pipeline: Human bias → Data → Algorithm → AI system → Decision making. [4] There are several ways behavior can appear from each of the steps before decision-making. For example, AI models are only as good as the data they are trained on. If the training data is biased, then the AI system will be biased. The algorithms used to train and run AI systems can also introduce bias if an algorithm heavily relies on a particular feature that is correlated with a protected characteristic–race, gender, age, religion, sexual orientation, disability, and national origin–leading to discriminatory outcomes.

3.4 APPLICATIONS and IMPACT

Social scientists have found that human judgments are often distorted by unconscious reasoning errors more frequently than people are aware of. [5] The reason employment interviews are still handled by humans with biases is due to the validity illusion. According to Kahneman and Tversky, who introduced the idea of cognitive bias into the modern era, confirmation bias causes people to overestimate their accuracy in making predictions, leading to an illusion of competence. [6] This is why using AI to omit irrelevant details is so appealing. However, if an AI is found to behave biased or unfairly when it comes to gender, race, economic status, or any given group, this may not only lead to a loss of trust with existing customers, but may also subject an organization to public and regulatory scrutiny, or harm those affected by the results.

A concern that organizations may have regarding transparency is that if faced with public backlash due to alleged bias or discrimination, they may have to halt the use of their ML systems or delay deployment, resulting in disruptions to their business operations. Amazon tried to implement an experimental AI recruiting tool. However, they decided to shut it down after discovering it discriminated against women. The tool would scrape the web to spot potential candidates and rate them. The algorithm learned to downgrade women’s CVs for technical jobs such as software development. In this case, the algorithm was trained on all CVs submitted to the company over ten years. Given the low proportion of women working in the company, the algorithm associated males with success. [7]

AI is gaining traction in the healthcare industry. Especially in analyzing medical images such as MRI scans or x-rays. According to a study published in PNAS in 2020, training data sets with gender imbalances in computer-aided diagnosis (CAD) systems resulted in reduced accuracy for the underrepresented group. [8] The JAMA Dermatology Network discovered disparities in skin cancer diagnosis across different skin colors. Models trained on light-skinned subjects were less accurate at identifying skin cancer in dark-skinned patients. Although dark-skinned individuals are generally at lower risk for skin cancer, it is often diagnosed at a more advanced stage in people of color which makes it harder to treat. The lack of diversity in AI training data results in models that struggle to diagnose skin cancer in patients with darker skin. [9]

In 2021, Twitter’s photo cropping function came under public scrutiny for allegedly being discriminatory towards black people and females as it favored white and male faces when cropping images. The initial thread that sparked this issue used an image of Mitch McConnell and Barack Obama. [10] Critics also accused Twitter’s photo cropping algorithm of emphasizing women’s chests and legs as more important features than their faces. In response to the backlash and criticism, Twitter investigated to determine the accuracy of the allegations and published technical reports detailing how the algorithms work to provide more transparency on the issue. Twitter also announced in a blog post that it was modifying the tool to eliminate the model biases that were identified. [11] However, there is controversy surrounding the exact factors that influenced Twitter’s photo cropping algorithm. While some initial reports suggested that the algorithm was discriminatory towards black people and women, further debate and experimentation in the comment section revealed conflicting information. Some sources claimed that the algorithm was primarily based on smile detection, as a similar image of Barack Obama with a bigger smile was chosen for the preview over the original photo of Mitch McConnell. Others found more factors such as overall contrast may have played a role. Nevertheless, Twitter faced significant backlash due to the initial assumptions about the algorithm’s bias. This underscores the importance of transparency in AI, as it can help to address misunderstandings and prevent negative consequences.

While the examples discussed in this section illustrate the challenges and limitations of current AI, explainable AI (XAI) offers a promising solution. XAI can provide human-understandable explanations of an AI system’s decision-making process, enabling users, companies, and researchers to understand how the system arrived at a particular output. This level of transparency and accountability can help address issues of bias and discrimination, making the technology more accurate, ethical, and fair in its decision-making.

Although the recognition and demand for XAI are increasing in various industries, bringing interpretability to AI is an incredibly challenging issue.

3.5 CHALLENGES

Due to the complex nature of modern machine learning algorithms, creating more explainable and interpretable AI poses significant challenges. For example, Deep Neural networks (DNN) are complex machine learning models that have achieved state-of-the-art performance in a wide range of applications including image recognition and natural language processing. However, DNNs have multi-layer nonlinear structures consisting of many hidden layers. Each layer can have numerous neurons and the model itself can have millions of parameters. The trade-off for this accuracy is a lack of interpretability. On the other hand, linear regressions are much simpler and easier to interpret. While they produce a formula that can be easily understood, as it assumes a linear relationship between the input variables and the output, it may not capture the complex interactions that a DNN can.

There is more to interpretability than just model design. Providing contextual interpretation and accurate explanations requires AI to understand the context in which the decision was made. This is challenging because context is complex, uncertain, and dynamic. Then, the ability to generate explanations both accurate and easy to understand by non-experts is another challenge. Different users may need different levels of information and intuition from the AI to be useful. This leads to XAI evaluation. From a sociological viewpoint, how does one evaluate the quality of an explanation? Luca Longo points to how this problem is framed such as the development of evaluation techniques from the field of data visualization. He specifically uses IKEA’s assembly diagrams as an example as they are meant to be interpretable by all “explainees”, not just an individual.

On the data side, the problem is its quality and bias embedded in it. As we have mentioned previously with technologies like Amazon’s recruiting tool, data biases can be introduced through various sources. This can lead to incorrect or incomplete results and explanations with XAI.

In addition to technical challenges, there are other issues to consider when applying XAI methods in various domains; some of which can have significant legal implications. For example, it is not yet clear whether explainability, in certain contexts, related to a model can be used to infer information about individuals. Techniques have been used in various forms of IT-Security, where anonymized or partially released sensitive data is analyzed to gain intelligence about the people involved. [12] Similar attacks have already been proposed and executed against machine learning models. [13] In addition, other legal challenges exist such as the right to be forgotten. [14]

There is a privacy problem. XAI methods can be used to reverse engineer personal information about individuals from seemingly safe data points. For example, even if the information is not explicitly stated, an XAI model could use an individual’s search history to infer their sexual orientation or political beliefs. This is where the right to be forgotten becomes relevant.

The right to be forgotten, in terms of the individual, is a legal concept that allows individuals to request the removal of their personal information from public databases, search engines, or other online platforms. It is considered an important privacy protection measure for individuals. While “data deletion” may seem to be a straightforward topic from the point of view of many regulators, this seemingly simple issue poses many practical problems in actual ML environments. In fact, “data deletion” requirements can be considered to border on the edge of impossibility.

Outputs of AI systems can contain personal information that is difficult or near-impossible to remove. For example, a model that predicts an individual’s risk of developing a certain disease may use personal health data that the individual wishes to keep private. If the individual requests that their data be removed from the model, it may be difficult or impossible to do so without retraining the entire model. This can be costly and time-consuming. XAI adds complexity to this issue as seen in a study where participants in ML datasets may have consented to be included on the assumption of anonymity. Several methods emerged in recent years that have proved capable of de-anonymizing personally identifiable information. Researchers from the National University of Singapore have concluded that “the more explainable AI becomes, the easier it will become to circumvent vital privacy features in machine learning systems.” [14, 22] They also found that it is possible to use explanations of similar models to ‘decode’ sensitive data in their corresponding non-explainable models.

3.6 CURRENT TECH and METHODS

3.6.1 Cyc

Cyc (/’saɪk’/) is a knowledge representation system that aims to provide common-sense knowledge to artificial intelligence systems. It consists of a large database of general knowledge and rules that allow AI to reason about the world in a more human-like way.

One of the main advantages of Cyc is its transparency. Because Cyc is based on explicitly defined knowledge and rules, it is possible to trace the reasoning process of an AI system that is based on Cyc. This makes it easier to understand how the AI system arrived at its decisions and to identify any errors or biases in its reasoning.

However, one of the limitations of Cyc is that it requires a significant amount of manual effort to maintain and update its knowledge base. This means that it may not be scalable to large and complex domains. Additionally, Cyc may not capture the full complexity and nuance of real-world situations, which can limit its usefulness in certain applications. [15]

3.6.2 DeepMind

Google’s DeepMind has developed several algorithms that are used for machine learning applications. However, these algorithms have been criticized for their lack of transparency.

To address these concerns, DeepMind has developed several tools and techniques to increase transparency in its algorithms. DeepLift is an algorithm that can be used to attribute the importance of input features to the output of a neural network. They have also developed an “explanation interface”; a graphical representation of the decision-making process. It allows users to understand how an algorithm made a particular decision. This can be useful in identifying errors or biases. [16]

3.6.3 GANs

Generative Adversarial Networks (GANs) are widely used in image and video generation tasks. However, the inner workings of GANs are often black boxes which makes it difficult to understand how they arrive at their outputs. To address this challenge, researchers have proposed several techniques for increasing the transparency of GANs.

One approach is to use attribution methods to identify the input features that are most important for generating the output. Attribution methods such as Gradient-weighted Class Activation Mapping (Grad-CAM) and Integrated Gradients can be used to make heatmaps and highlight regions of the input image that contribute the most to the output generated by the GAN. [17]

In addition, researchers have also proposed using adversarial attacks to evaluate the transparency of GANs. Adversarial attacks involve intentionally perturbing–adding disturbances–the input to a neural network to generate an output that is different from the expected output. By evaluating how well a GAN can resist these attacks, researchers can gain insights into the robustness and transparency of the network. [18]

Overall, these techniques for increasing the transparency of GANs can help researchers better understand and improve the performance of these models by identifying potential biases and limitations in the network.

3.6.4 MODEL-AGNOSTIC METHODS

Outside of specific techniques and technologies, several important model-agnostic interpretability methods can help interpret the results of AI and ML models.

Partial Dependence Plots (PDP), introduced by Leo Breiman and Adele Cutler in 1994, is a global method that is a type of visual explanation tool used to understand the relationship between a feature and a target variable. They show how the predicted outcome of the model changes as the value of a single feature changes. PDP is intuitive but does not account for heterogeneous effects–the effect of one variable on an outcome can be different for different groups. Individual Conduction Expectation (ICE) solves this problem. ICE is a local method that provides the effect of changing a feature on the model for each instance where PDP is the average effect. [19]

Local Interpretable Model-agnostic Explanations (LIME) work by approximating the original model with an interpretable model. It creates perturbations and trains interpretable models on these. LIME tries to understand how the predictions change when variations are added to the data. For example, applying LIME to an image classifier would divide an image into components, and the highest positive weights assigned to portions are presented as the explanation. [19]

Shapely Additive Explanations (SHAP) are based on game theory. A prediction can be explained by assuming each feature or SHAP value of an instance is a “player”. Each SHAP value is estimated by adding and removing the player from subsets with the other players which leads to a full interpretation. [19]

There are three important considerations when deciding which method to use. The first is how much of the model you need to understand. This will help determine whether to use a global or local method. Second, if there is a time limitation, a method like SHAP is extremely complex and exhaustive and may not be appropriate. Finally, knowing the user’s experience can help decide whether more sophisticated or simple answers are suitable.

3.6.5 MODEL EVALUATION

While model interpretation techniques can show how machine learning models work, evaluating their interpretability is also important. Model evaluation can help determine whether a model’s outputs are reliable, accurate, and understandable. There are three main categories of model evaluation techniques: application-grounded evaluation, human-grounded evaluation, and functionally-grounded evaluation.

Application-grounded evaluation involves evaluating a model’s performance on specific tasks and understanding how its outputs relate to the real-world application. This approach can help identify potential biases or limitations in the model and suggest improvements for future iterations. To perform application-grounded evaluation, researchers typically define a set of tasks or applications that the model is intended to solve. They then evaluate the model’s performance on these tasks using standard metrics such as accuracy, precision, recall, and F1 score. By analyzing the model’s performance on these tasks, researchers can gain insights into its strengths and weaknesses, identifying areas for improvement. One key advantage of application-grounded evaluation is that it provides a clear and practical way to assess the usefulness of a machine learning model. By focusing on real-world applications, this approach helps to ensure that the model is solving a functional problem and that its outputs are relevant and meaningful. However, there are also some limitations. For example, this approach may only sometimes capture the full complexity of real-world situations. This can limit its usefulness in specific applications. Additionally, it may be difficult to define a set of tasks that fully capture the scope of a machine learning model’s capabilities. [19]

Human-grounded evaluation involves using human experts to evaluate the interpretability of a model. This type of evaluation is particularly useful when the interpretability of a model is important for decision-making, such as in medical diagnosis or hiring decisions. It can take different forms depending on the specific problem being addressed. One approach is to have domain experts review the model’s outputs and provide feedback on its interpretability. This feedback can be used to improve its transparency and help the experts understand how it is making its predictions. Another approach is to conduct experiments to test how well people can understand and use the model’s outputs. For example, researchers might ask study participants to make decisions based on the model’s predictions and compare their decisions to those made using other sources of information such as expert knowledge or basic Google research. This can provide insights into how well the model’s outputs align with human decision-making and whether improvements to the model’s interpretability would be useful. [19]

Functionally-grounded evaluation, unlike application- or human-grounded evaluation, focuses on evaluating the effectiveness of an interpretability method in terms of its impact on the model’s overall performance. This type of evaluation requires comparing the performance of the original model without any interpretability method to the performance of the model with the interpretability method applied. The effectiveness of the interpretability method is measured by the degree to which it improves the performance of the model, while still maintaining the desired level of interpretability.

One common approach to functionally grounded evaluation is to use a metric that takes into account both performance and interpretability, such as the accuracy-interpretability trade-off. This approach involves measuring the accuracy of the model and comparing it to its level of interpretability. The interpretability can be measured using a metric such as the number of features used in the model or the complexity of the model. [20]

Another approach to functionally grounded evaluation is to use a perturbation-based method, where the interpretability method is evaluated by the degree to which it affects the model’s output when a perturbation or disturbance is applied to the input data. The change can be random noise or the removal or modification of certain features in the input.

Functionally grounded evaluation is an important aspect of evaluating interpretability methods as it provides insight into how interpretability can be achieved without compromising the model’s performance. [19]

3.6.6 PRIVACY TECHNIQUES

Privacy-preserving techniques like differential privacy, federated learning, and homomorphic encryption can help to protect sensitive data. Differential privacy is a mathematical framework that ensures that the results of an analysis or model do not reveal anything about the individual data points that were used to create it. It works by adding random noise to the data before it is analyzed, which makes it more difficult for an attacker to identify any particular individual’s information. [14]

Federated learning started as a way to help Android users solve the problem of updating their models locally. It can be applied to various fields of machine learning. The focus turned to the problem of learning model parameters when data is distributed across multiple nodes without having to send raw data to a central node. Federated learning has been used in Google’s Gboard system to predict keyboard input to protect privacy. It is also very useful in the medical field as patients’ medical data is sensitive.

Homomorphic encryption has become commonly used with distributed ML as it enables distributed parties to jointly compute an arbitrary functionality without revealing their private inputs and outputs. By incorporating these techniques into XAI designs, privacy can be protected while still allowing for the creation of powerful models and insights. [21]

4 CONCLUSION

In this paper, we have discussed the importance of transparency in AI and explored various techniques for interpreting AI systems and ML models. Interpretable and explainable models are crucial for building trust, detecting bias, and ensuring the ethical use of AI.

The hope is for future AI systems to be designed with high accuracy and explainability in mind. All stages from data collection to algorithm development should have some degree of transparency depending on the importance of the tool in its environment. Tools, such as ChatGPT, may be useful to embed and deliver highly interpretable natural language explanations to those of any expertise. To ensure that explanations are relevant, it is also important to define the scope of explanations and involve humans in the loop. Research for generating explanations should also investigate how they affect and influence people’s understanding and actions.

We hope that the findings of this study will help further the understanding of the XAI’s role in ensuring the responsible and ethical use of AI.

REFERENCES

[1] “Artificial Intelligence Market Size & Share Analysis Report 2030.” https://www.grandviewresearch.com/industry-analysis/artificial-intelligence-ai-market (accessed Mar. 02, 2023).

[2] “Explainable Artificial Intelligence.” https://www.darpa.mil/program/explainable-artificial-intelligence (accessed Mar. 02, 2023).

[3] J. Greene, F. Rossi, J. Tasioulas, K. Venable, and B. Williams, “Embedding Ethical Principles in Collective Decision Support Systems,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 30, no. 1, Art. no. 1, Mar. 2016, doi: 10.1609/aaai.v30i1.9804. (accessed Mar. 02, 2023).

[4] M. Vasconcelos, C. Cardonha, and B. Gonçalves, “Modeling Epistemological Principles for Bias Mitigation in AI Systems: An Illustration in Hiring Decisions,” in Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, Dec. 2018, pp. 323–329. doi: 10.1145/3278721.3278751. (accessed Mar. 02, 2023).

[5] K. A. Houser, “Can AI Solve the Diversity Problem in the Tech Industry? Mitigating Noise and Bias in Employment Decision-Making,” 2019. (accessed Mar. 02, 2023).

[6] D. Kahneman and A. Tversky, “On the Psychology of Prediction,” Psychological Review, vol. 80, no. 4, pp. 237–251, 1973, doi: 10.1037/h0034747. (accessed Mar. 02, 2023).

[7] “Amazon’s sexist hiring algorithm could still be better than a human,” Aug. 19, 2022. https://www.imd.org/research-knowledge/articles/amazons-sexist-hiring-algorithm-could-still-be-better-than-a-human/ (accessed Mar. 02, 2023).

[8] A. J. Larrazabal, N. Nieto, V. Peterson, D. H. Milone, and E. Ferrante, “Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis,” Proceedings of the National Academy of Sciences, vol. 117, no. 23, pp. 12592–12594, Jun. 2020, doi: 10.1073/pnas.1919012117. (accessed Mar. 02, 2023).

[9] “Machine Learning and Health Care Disparities in Dermatology | Dermatology | JAMA Dermatology | JAMA Network.” https://jamanetwork.com/journals/jamadermatology/article-abstract/2688587 (accessed Mar. 02, 2023).

[10] “Tweet / Twitter,” Twitter. https://twitter.com/tamaleaver/status/1307466077861896193 (accessed Mar. 02, 2023).

[11] “Overcoming AI’s Transparency Paradox.” https://www.forbes.com/sites/aparnadhinakaran/2021/09/10/overcoming-ais-transparency-paradox/?sh=3cd840204b77 (accessed Mar. 02, 2023).

[12] L. Longo, R. Goebel, F. Lecue, P. Kieseberg, and A. Holzinger, “Explainable Artificial Intelligence: Concepts, Applications, Research Challenges and Visions,” in Machine Learning and Knowledge Extraction, Cham, 2020, pp. 1–16. doi: 10.1007/978–3–030–57321–8_1.

[13] L. Rizzo, L. Majnaric, and L. Longo, “A Comparative Study of Defeasible Argumentation and Non-monotonic Fuzzy Reasoning for Elderly Survival Prediction Using Biomarkers,” in AI*IA 2018 — Advances in Artificial Intelligence, Cham, 2018, pp. 197–209. doi: 10.1007/978–3–030–03840–3_15. (accessed Mar. 02, 2023).

[14] E. F. Villaronga, P. Kieseberg, and T. Li, “Humans forget, machines remember: Artificial intelligence and the Right to Be Forgotten,” Computer Law & Security Review, vol. 34, no. 2, pp. 304–313, Apr. 2018, doi: 10.1016/j.clsr.2017.08.007. (accessed Mar. 02, 2023).

[15] Cyc. Cycorp. [Online] Available: https://www.cyc.com/wp-content/uploads/2019/09/Cyc-Technology-Overview.pdf (accessed Mar. 03. 2023).

[16] M. D. Zeiler and R. Fergus, “Visualizing and Understanding Convolutional Networks,” in Computer Vision — ECCV 2014, vol. 8689, D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, Eds. Cham: Springer International Publishing, 2014, pp. 818–833. doi: 10.1007/978–3–319–10590–1_53. (accessed Mar. 04, 2023).

[17] S.-M. Shih, P.-J. Tien, and Z. Karnin, “GANMEX: One-vs-One Attributions using GAN-based Model Explainability,” in Proceedings of the 38th International Conference on Machine Learning, Jul. 2021, pp. 9592–9602. Accessed: Mar. 02, 2023. [Online]. Available: https://proceedings.mlr.press/v139/shih21a.html (accessed Mar. 04, 2023).

[18] “Exploring generative adversarial networks and adversarial training — ScienceDirect.” https://www.sciencedirect.com/science/article/pii/S2666307422000080 (accessed Mar. 02, 2023).

[19] F. Doshi-Velez and B. Kim, “Towards A Rigorous Science of Interpretable Machine Learning.” arXiv, Mar. 02, 2017. Accessed: Mar. 04, 2023. [Online]. Available: http://arxiv.org/abs/1702.08608

[20] G. Myrianthous, “Understanding The Accuracy-Interpretability Trade-Off,” Medium, Oct. 06, 2021. https://towardsdatascience.com/accuracy-interpretability-trade-off-8d055ed2e445 (accessed Mar. 04, 2023).

[21] H. Fang and Q. Qian, “Privacy Preserving Machine Learning with Homomorphic Encryption and Federated Learning,” Future Internet, vol. 13, no. 4, Art. no. 4, Apr. 2021, doi: 10.3390/fi13040094.

[22] “Explainable AI May Surrender Confidential Data More Easily — Unite.AI.” https://www.unite.ai/explainable-ai-may-surrender-confidential-data-more-easily/ (accessed Mar. 08, 2023).