Human Machine Interaction and Security in the era of modern Machine Learning

Abstract

It is realistic to describe Artificial Intelligence (AI) as the most important of emerging technologies because of its increasing dominance in almost every field of modern life and the crucial role it plays in boosting high-tech multidisciplinary developments integrated in steady innovations. The implementation of AI-based solutions for real world problems helps to create new insights into old problems and to produce unique knowledge about intractable problems which are too complex to be efficiently solved with conventional methods. Biomedical data analysis, computer-assisted drug discovery, pandemic predictions and preparedness are only but a few examples of applied research areas that use machine learning as a pivotal data evaluation tool. Such tools process enormous amounts of data trying to discover causal relations and risk factors and predict outcomes that for example can change the course of diseases. The growing number of remarkable achievements delivered by modern machine learning algorithms in the last years raises enthusiasm for all those things that AI can do. The value of the global artificial intelligence market was calculated at USD 136.55 billion in 2022 and is estimated to expand at an annual growth rate of 37.3% from 2023 to 2030. Novel machine-learning applications in finance, national security, health, criminal justice, transportation, smart cities etc. justify the forecast that AI will have a disruptive impact on economies, societies and governance. The traditional rule-based or expert systems, known in computer science since decades implement factual, widely accepted knowledge and heuristic of human experts and they operate by practically imitating the decision making process and reasoning functionalities of professionals. In contrast, modern statistical machine learning systems discover their own rules based on examples on the basis of vast amounts of training data introduced to them. Unfortunately the predictions of these systems are generally not understandable by humans and quite often they are neither definite or unique. Raising the accuracy of the algorithms doesn't improve the situation. Various multi-state initiatives and business programs have been already launched and are in progress to develop technical and ethical criteria for reliable and trustworthy artificial intelligence. Considering the complexity of famous leading machine learning models (up to hundreds of billion parameters) and the influence they can exercise for example by creating text and news and also fake news, generate technical articles, identify human emotions, identify illness etc. it is necessary to expand the definition of HMI (Human Machine Interface) and invent new security concepts associated with it. The definition of HMI has to be extended to account for real-time procedural interactions of humans with algorithms and machines, for instance when faces, body movement patterns, thoughts, emotions and so on are considered to become available for classification both with or without the person's consent. The focus of this work will be set upon contemporary technical shortcomings of machine learning systems that render the security of a plethora of new kinds of human machine interactions as inadequate. Examples will be given with the purpose to raise awareness about underestimated risks.