Adversarial Attacks and Defences: A Survey
Source: Dev.to
Overview
Today many apps use deep learning to perform complex tasks quickly, from image analysis to voice recognition. However, tiny, almost invisible changes to the input can cause a model to give the wrong answer—these are known as adversarial attacks. Such attacks appear harmless but can lead to misclassification, potentially breaking services or compromising security.
Researchers aim to improve AI robustness, yet universal fixes are rare. Some methods work in specific scenarios, while others fail when the adversarial perturbation is slightly altered. The result is that powerful AI systems remain useful but fragile, and attackers could exploit these weaknesses.
Designers need to test models thoroughly, monitor for suspicious inputs, and implement multiple layers of protection. Although it is impossible to prevent every adversarial trick, simple checks and careful design can reduce the likelihood of unexpected failures.
A small change in an image can alter a significant decision, so awareness and vigilance are essential.