Reducing risks when using machine learning in diagnosis of bronchopulmonary diseases

The article is about issues of risk reduction when using software solutions based on machine learning methods for classifying chest x-rays on the example of chest x-ray in the diagnosis of bronchopulmonary diseases. A problem statement is formulated to reduce the risk of misdiagnosis by using of methods to counter malicious attacks. The machine learning methods of classification problem, the most dangerous attacks that reduce the recognition efficiency, and measures to counter attacks to reduce risks are identified based on experimental data. These methods were used when experimental studies. Defensive distillation, filtration, unlearning, pruning were used as countermeasures. The results obtained allow us to state that these methods can be used for other images as well. The results of experimental studies made it possible to formulate recommendations as rules, including combinations of recognition methods, attacks, and countermeasures to reduce the risk of misdiagnosis.