An Annealing Mechanism for Adversarial Training Acceleration

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks are susceptible to adversarial attacks, degrading performance.
• Adversarial training enhances robustness but incurs high computational costs.

Purpose of the Study:

• To reduce the computational overhead of adversarial training.
• To accelerate the process of making deep neural networks more robust.

Main Methods:

• Proposed an annealing mechanism for adversarial training acceleration (Amata).
• Amata is provably convergent and motivated by optimal control theory.
• Amata can be combined with existing acceleration methods.

Main Results:

• Amata achieves similar or better robustness with 1/3-1/2 the computational time.
• Demonstrated effectiveness on standard datasets.
• Amata integration with other methods further reduces computation on large-scale problems.

Conclusions:

• Amata offers a computationally efficient solution for adversarial training.
• The mechanism enhances robustness without prohibitive training times.
• Amata is a versatile tool for improving deep learning security.