Resistance of deep learning models to adversarial attacks

Ovaj rad istražuje suparničke napade na duboke neuronske mreže korištene za probleme računalnog vida. Suparnički napadi mijenjaju ulazne podatke kako bi model strojnog učenja donio pogrešne zaključke. Karakteristika ovih napada je da su promjene na ulaznim podacima gotovo neprimjetne ljudima dok značajno utječu na model. Uvodni dio rada detaljno objašnjava duboke neuronske mreže, njihov rad te osjetljivost na suparničke napade. Kasnije u radu, detaljno su objašnjeni različiti načini kreiranja suparničkih napada, uključujući algoritme korištene za njihovo generiranje. Opisane su i metode obrana kojima štitimo duboke neuronske mreže, kako bi postale robusnije. Provedeni su eksperimenti kojima je utvrđena učinkovitost suparničkih napada i obrana od njih, a rad je zaključen objašnjenjem dobivenih rezultata.This thesis investigates adversarial attacks on deep neural networks used for computer vision problems. Adversarial attacks modify input data in order to cause the machine learning model to make incorrect conclusions. The characteristic of these attacks is that the changes made to the input data are nearly imperceptible to humans, while significantly impacting machine learning model. The introduction of the thesis provides a detailed explanation of deep neural networks, their functioning and their vulnerability to adversarial attacks. Later in the thesis, various methods of creating adversarial attacks are explained in detail, including the specific algorithms used for generating them. Defense methods are described in order to protect deep neural networks and make them more robust. Various experiments have been conducted to evaluate the effectiveness of adversarial attacks and defenses against them, and the thesis is concluded by interpreting the obtained result