Super-Resolution

Super-Resolution enhances image quality and mitigates adversarial attacks on convolutional neural networks by applying wavelet denoising and deep image super-resolution to map off-the-manifold adversarial samples back onto the natural image manifold.

Key Features:

• Adversarial Defense: Mitigates adversarial noise patterns and restores CNN classification accuracy by enhancing images through deep restoration networks.
• Wavelet Denoising: Applies wavelet denoising as a preprocessing step to reduce adversarial perturbations.
• Image Super-Resolution: Uses deep image super-resolution to improve image clarity and map perturbed samples toward the natural image manifold.
• Model-Agnostic Approach: Operates without modifying the target classifier and does not require a separate adversarial-detection mechanism.
• Complementary and Efficient: Complements existing defense mechanisms and functions without additional training or parameter optimization.
• Broad Applicability: Demonstrates superior performance across popular adversarial attack algorithms.

Scientific Applications:

• Medical imaging: Preserves image integrity and classification performance in clinical imaging pipelines subject to adversarial perturbations.
• Surveillance: Enhances robustness of surveillance imaging systems against adversarial manipulation.
• Autonomous vehicles: Improves reliability of vision-based perception in autonomous driving by reducing adversarial influence on sensor images.
• Computer vision systems: Maintains high performance of general computer vision models on both clean and adversarially perturbed images.

Methodology:

Preprocessing combines wavelet denoising and deep image super-resolution via deep image restoration networks to map off-the-manifold adversarial samples back onto the natural image manifold; the approach is computationally efficient and requires no additional training or parameter tuning.