Research Papers

Intentional and accidental harms arising from the use of AI have impacted the health, safety and rights of individuals. While regulatory frameworks are being developed, there remains a lack of consensus on methods necessary to deliver safe AI. The potential for explainable AI (XAI) to contribute to the effectiveness of the regulation of AI is being increasingly examined. Regulation must include methods to ensure compliance on an ongoing basis, though there is an absence of practical proposals on how to achieve this. For XAI to be successfully incorporated into a regulatory system, the individuals who are engaged in interpreting/explaining the model to stakeholders should be sufficiently qualified for the role. Statutory professionals are prevalent in domains in which harm can be done to the health, safety and rights of individuals. The most obvious examples are doctors, engineers and lawyers. Those professionals are required to exercise skill and judgement and to defend their decision making process in the event of harm occurring. We propose that a statutory profession framework be introduced as a necessary part of the AI regulatory framework for compliance and monitoring purposes. We will refer to this new statutory professional as an AI Architect (AIA). This AIA would be responsible to ensure the risk of harm is minimised and accountable in the event that harms occur. The AIA would also be relied on to provide appropriate interpretations/explanations of XAI models to stakeholders. Further, in order to satisfy themselves that the models have been developed in a satisfactory manner, the AIA would require models to have appropriate transparency. Therefore it is likely that the introduction of an AIA system would lead to an increase in the use of XAI to enable AIA to discharge their professional obligations.

Explainability poses a major challenge to artificial intelligence (AI) techniques. Current studies on explainable AI (XAI) lack the efficiency of extracting global knowledge about the learning task, thus suffer deficiencies such as imprecise saliency, context-aware absence and vague meaning. In this paper, we propose the class association embedding (CAE) approach to address these issues. We employ an encoder-decoder architecture to embed sample features and separate them into class-related and individual-related style vectors simultaneously. Recombining the individual-style code of a given sample with the class-style code of another leads to a synthetic sample with preserved individual characters but changed class assignment, following a cyclic adversarial learning strategy. Class association embedding distills the global class-related features of all instances into a unified domain with well separation between classes. The transition rules between different classes can be then extracted and further employed to individual instances. We then propose an active XAI framework which manipulates the class-style vector of a certain sample along guided paths towards the counter-classes, resulting in a series of counter-example synthetic samples with identical individual characters. Comparing these counterfactual samples with the original ones provides a global, intuitive illustration to the nature of the classification tasks. We adopt the framework on medical image classification tasks, which show that more precise saliency maps with powerful context-aware representation can be achieved compared with existing methods. Moreover, the disease pathology can be directly visualized via traversing the paths in the class-style space.

39 seconds. That is the timelapse between two consecutive cyber attacks as of 2023. Meaning that by the time you are done reading this abstract, about 1 or 2 additional cyber attacks would have occurred somewhere in the world. In this context of highly increased frequency of cyber threats, Security Operation Centers (SOC) and Computer Emergency Response Teams (CERT) can be overwhelmed. In order to relieve the cybersecurity teams in their investigative effort and help them focus on more added-value tasks, machine learning approaches and methods started to emerge. This paper introduces a novel method, IsoEx, for detecting anomalous and potentially problematic command lines during the investigation of contaminated devices. IsoEx is built around a set of features that leverages the log structure of the command line, as well as its parent/child relationship, to achieve a greater accuracy than traditional methods. To detect anomalies, IsoEx resorts to an unsupervised anomaly detection technique that is both highly sensitive and lightweight. A key contribution of the paper is its emphasis on interpretability, achieved through the features themselves and the application of eXplainable Artificial Intelligence (XAI) techniques and visualizations. This is critical to ensure the adoption of the method by SOC and CERT teams, as the paper argues that the current literature on machine learning for log investigation has not adequately addressed the issue of explainability. This method was proven efficient in a real-life environment as it was built to support a companyś SOC and CERT

Quantifying the Intrinsic Usefulness of Attributional Explanations for Graph Neural Networks with Artificial Simulatability Studies

Jonas Teufel, Luca Torresi, Pascal Friederich

Despite the increasing relevance of explainable AI, assessing the quality of explanations remains a challenging issue. Due to the high costs associated with human-subject experiments, various proxy metrics are often used to approximately quantify explanation quality. Generally, one possible interpretation of the quality of an explanation is its inherent value for teaching a related concept to a student. In this work, we extend artificial simulatability studies to the domain of graph neural networks. Instead of costly human trials, we use explanation-supervisable graph neural networks to perform simulatability studies to quantify the inherent usefulness of attributional graph explanations. We perform an extensive ablation study to investigate the conditions under which the proposed analyses are most meaningful. We additionally validate our methods applicability on real-world graph classification and regression datasets. We find that relevant explanations can significantly boost the sample efficiency of graph neural networks and analyze the robustness towards noise and bias in the explanations. We believe that the notion of usefulness obtained from our proposed simulatability analysis provides a dimension of explanation quality that is largely orthogonal to the common practice of faithfulness and has great potential to expand the toolbox of explanation quality assessments, specifically for graph explanations.

The work covers the development and explainability of machine learning models for predicting political leanings through parliamentary transcriptions. We concentrate on the Slovenian parliament and the heated debate on the European migrant crisis, with transcriptions from 2014 to 2020. We develop both classical machine learning and transformer language models to predict the left- or right-leaning of parliamentarians based on their given speeches on the topic of migrants. With both types of models showing great predictive success, we continue with explaining their decisions. Using explainability techniques, we identify keywords and phrases that have the strongest influence in predicting political leanings on the topic, with left-leaning parliamentarians using concepts such as people and unity and speak about refugees, and right-leaning parliamentarians using concepts such as nationality and focus more on illegal migrants. This research is an example that understanding the reasoning behind predictions can not just be beneficial for AI engineers to improve their models, but it can also be helpful as a tool in the qualitative analysis steps in interdisciplinary research.

Carbon offset programs are critical in the fight against climate change. One emerging threat to the long-term stability and viability of forest carbon offset projects is wildfires, which can release large amounts of carbon and limit the efficacy of associated offsetting credits. However, analysis of wildfire risk to forest carbon projects is challenging because existing models for forecasting long-term fire risk are limited in predictive accuracy. Therefore, we propose an explainable artificial intelligence (XAI) model trained on 7 million global satellite wildfire observations. Validation results suggest substantial potential for high resolution, enhanced accuracy projections of global wildfire risk, and the model outperforms the U.S. National Center for Atmospheric Research's leading fire model. Applied to a collection of 190 global forest carbon projects, we find that fire exposure is projected to increase 55% [37-76%] by 2080 under a mid-range scenario (SSP2-4.5). Our results indicate the large wildfire carbon project damages seen in the past decade are likely to become more frequent as forests become hotter and drier. In response, we hope the model can support wildfire managers, policymakers, and carbon market analysts to preemptively quantify and mitigate long-term permanence risks to forest carbon projects.

Recommender system is the most successful commercial technology in the past decade. Technical mammoth such as Temu, TikTok and Amazon utilize the technology to generate enormous revenues each year. Although there have been enough research literature on accuracy enhancement of the technology, explainable AI is still a new idea to the field. In 2022, the author of this paper provides a geometric interpretation of the matrix factorization-based methods and uses geometric approximation to solve the recommendation problem. We continue the research in this direction in this paper, and visualize the inner structure of the parameter space of matrix factorization technologies. We show that the parameters of matrix factorization methods are distributed within a hyper-ball. After further analysis, we prove that the distribution of the parameters is not multivariate normal.

The prediction of age is a challenging task with various practical applications in high-impact fields like the healthcare domain or criminology. Despite the growing number of models and their increasing performance, we still know little about how these models work. Numerous examples of failures of AI systems show that performance alone is insufficient, thus, new methods are needed to explore and explain the reasons for the model's predictions. In this paper, we investigate the use of Explainable Artificial Intelligence (XAI) for age prediction focusing on two specific modalities, EEG signal and lung X-rays. We share predictive models for age to facilitate further research on new techniques to explain models for these modalities.

Recent advancements in AI have coincided with ever-increasing efforts in the research community to investigate, classify and evaluate various methods aimed at making AI models explainable. However, most of existing attempts present a method-centric view of eXplainable AI (XAI) which is typically meaningful only for domain experts. There is an apparent lack of a robust qualitative and quantitative performance framework that evaluates the suitability of explanations for different types of users. We survey relevant efforts, and then, propose a unified, inclusive and user-centred taxonomy for XAI based on the principles of General System's Theory, which serves us as a basis for evaluating the appropriateness of XAI approaches for all user types, including both developers and end users.

Deep learning is becoming increasingly adopted in business and industry due to its ability to transform large quantities of data into high-performing models. These models, however, are generally regarded as black boxes, which, in spite of their performance, could prevent their use. In this context, the field of eXplainable AI attempts to develop techniques that temper the impenetrable nature of the models and promote a level of understanding of their behavior. Here we present our contribution to XAI methods in the form of a framework that we term SpecXAI, which is based on the spectral characterization of the entire network. We show how this framework can be used to not only understand the network but also manipulate it into a linear interpretable symbolic representation.

This paper's main goal is to provide an attacker's point of view on data poisoning assaults that use label-flipping during the training phase of systems that use electroencephalogram (EEG) signals to evaluate human emotion. To attack different machine learning classifiers such as Adaptive Boosting (AdaBoost) and Random Forest dedicated to the classification of 4 different human emotions using EEG signals, this paper proposes two scenarios of label-flipping methods. The results of the studies show that the proposed data poison attacksm based on label-flipping are successful regardless of the model, but different models show different degrees of resistance to the assaults. In addition, numerous Explainable Artificial Intelligence (XAI) techniques are used to explain the data poison attacks on EEG signal-based human emotion evaluation systems.

Training dataset biases are by far the most scrutinized factors when explaining algorithmic biases of neural networks. In contrast, hyperparameters related to the neural network architecture have largely been ignored even though different network parameterizations are known to induce different implicit biases over learned features. For example, convolutional kernel size is known to affect the frequency content of features learned in CNNs. In this work, we present a causal framework for linking an architectural hyperparameter to out-of-distribution algorithmic bias. Our framework is experimental, in that we train several versions of a network with an intervention to a specific hyperparameter, and measure the resulting causal effect of this choice on performance bias when a particular out-of-distribution image perturbation is applied. In our experiments, we focused on measuring the causal relationship between convolutional kernel size and face analysis classification bias across different subpopulations (race/gender), with respect to high-frequency image details. We show that modifying kernel size, even in one layer of a CNN, changes the frequency content of learned features significantly across data subgroups leading to biased generalization performance even in the presence of a balanced dataset.