Deep learning is a crucial aspect of machine learning, but it also makes these techniques vulnerable to adversarial examples, which can be seen in a variety of applications. These examples can even be targeted at humans, leading to the creation of false media, such as deepfakes, which are often used to shape public opinion and damage the reputation of public figures. This article will explore the concept of adversarial examples, which are comprised of perturbations added to clean images or videos, and their ability to deceive DL algorithms. The proposed approach achieved a precision value of accuracy of 76.2% on the DFDC dataset.

翻译：深度学习是机器学习的关键组成部分，但也使这些技术容易受到对抗样本的攻击，这一问题在多种应用中均有体现。这些对抗样本甚至可能针对人类，导致虚假媒体（例如深度伪造）的生成，此类内容常被用于塑造公众舆论或损害公众人物的声誉。本文将探讨对抗样本的概念——即向干净图像或视频中添加扰动所构成的样本——及其对深度学习算法的欺骗能力。所提出的方法在DFDC数据集上达到了76.2%的准确率精度值。