Methods for Defense Against Backdoor Attacks Upon Neural Networks Using Data Augmentation

Dan je pregled sigurnosnih opasnosti koji prijete modelima dubokog učenja uz fokus na utjecaj primjene augmentacije podataka na sigurnost, metrike i performanse modela dubokog učenja. Formalno su opisani backdoor napadi na modele te je dan pregled obrambenih metoda. Za provedbu eksperimenata, razvijena je programska implementacija procesa učenja inficiranih modela te provođenja obrane u obliku augmentacije podataka točno definiranim strategijama transformacija slika u Keras i PyTorch frameworcima te je pokazan utjecaj augmentacije podataka na osiguravanje modela od umetnutoga backdoora.An overview of security threats to deep learning models is given, with a focus on the impact of applying data augmentation on the security, metrics, and performance of deep learning models. Backdoor attacks on models are formally described and an overview of defense methods is given. To carry out the experiments, a software implementation of the process of training infected models and a defense method in the form of data augmentation with precisely defined image transformation policies was developed in Keras and PyTorch frameworks, and the impact of data augmentation on the model security from an inserted backdoor was demonstrated