Hands-On Artificial Intelligence for Cybersecurity

Defending against adversarial attacks

Following the growing diffusion of adversarial attacks, many attempts have been made to provide adequate defense measures, based mainly on the following methods:

• Statistical-based detection defense: This method tries to detect the presence of adversarial examples by exploiting statistical tests and outlier detection. It assumes that the statistical distributions characterizing the real examples and the adversarial examples are fundamentally distinct from one another. However, the effectiveness of the C and W attack methodology shows that this assumption is not at all obvious or reliable.
• Gradient masking defense: We have seen how adversarial attacks exploit the backpropagation optimization strategy adopted ...