Push-to-Corner Preprocessing for Adversarial Defense
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks are vulnerable to adversarial attacks, which can lead to incorrect predictions and malicious activities, posing a security concern in deep learning systems, and existing defense mechanisms like model distillation and adversarial training may not provide sufficient robustness against sophisticated attacks.
Innovation Solution
The implementation of a push-to-corner preprocessing method that uses a saturation function to push input data towards the corners of its domain, combined with adversarial training, to generate adversarial attack data and enhance the robustness of neural networks against such attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If adversarial training is applied to improve robustness against adversarial attacks, then reliability improves, but accuracy on unperturbed data deteriorates
Solution Approach 1:
The patent applies push-to-corner preprocessing as a preliminary action before the main classification task. By transforming input data to push corner values toward the center of their domains, the method prepares the data in advance to reduce vulnerability to adversarial attacks, thereby improving robustness while maintaining acceptable accuracy on clean data
Solution Approach 2:
The patent modifies the input data parameters through push-to-corner preprocessing, which transforms the distribution of input values by pushing corner values toward the center. This parameter transformation helps the neural network become more robust to adversarial perturbations while preserving the essential information needed for accurate classification
2Reliability
If push-to-corner preprocessing is applied to enhance security against adversarial attacks, then reliability improves, but manufacturing precision deteriorates
Solution Approach 1:
The patent converts the harmful effect of push-to-corner preprocessing on perturbed data accuracy into a benefit by combining it with adversarial training. The preprocessing step that might degrade perturbed data accuracy is transformed into a defensive mechanism that, when paired with adversarial examples during training, ultimately improves robustness against real adversarial attacks
Data Source
AI summary
A platform for training deep neural networks using push-to-corner preprocessing and adversarial training. A training engine adds a preprocessing layer before the input data is fed into a deep neural network at the input layer, for pushing the input data further to the corner of its domain.