Encoder Adversarial Pretraining Using Similar Target Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models are vulnerable to adversarial attacks, and existing unsupervised adversarial learning methods often generate inefficient adversarial examples, particularly in non-contrast self-supervised learning models, leading to reduced model accuracy.
Innovation Solution
An adversarial self-supervised learning method that selects a target image similar to the original image and generates a noise image to train the encoder, using transformation and loss functions to enhance robustness against adversarial attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If unsupervised adversarial learning is implemented by maximizing self-supervised learning loss for each image, then the model can be trained without label information, but the generated adversarial examples become inefficient and model accuracy decreases
Solution Approach 1:
The patent introduces a target image as an intermediary element in the adversarial learning process. Instead of maximizing self-supervised loss for each image independently, the method uses a target image (selected from an image group) as a reference to generate adversarial examples. This intermediary target image guides the adversarial perturbation process, ensuring that the generated noise images effectively challenge the model while maintaining training efficiency and improving accuracy.
2Reliability
If noise is generated in a direction in which a loss function is incorrect during learning, then the deep learning model can be trained to lower the loss function, but the model becomes vulnerable to adversarial attacks with prediction performance dropping to 0%
Solution Approach 1:
The patent converts the harmful effect of adversarial noise into a beneficial training mechanism. By generating noise images using the original image and a target image as inputs, and training the encoder to minimize the difference between their feature representations, the method transforms the previously harmful adversarial noise into a useful training signal. This process makes the model robust to adversarial attacks while maintaining the ability to reduce loss function during training.
3Reliability
If a deep learning model is trained with clean images, then the model achieves high performance initially, but it reacts sensitively to very small amounts of noise and fails to produce original performance
Solution Approach 1:
The patent applies preliminary action by pre-training the encoder on clean images to achieve high initial performance, then subsequently training it on generated noise images to build robustness. The encoder is first trained to accurately represent clean image data, and then exposed to adversarial noise images where it learns to distinguish between legitimate variations and adversarial perturbations. This two-stage approach preserves initial performance while adding noise resistance.
Data Source
AI summary
In accordance with an aspect of the present disclosure, there is provided an adversarial self-supervised learning method for an encoder. The method comprises selecting a target image of an original image for training from an image group included in a previously collected dataset; generating a noise image by combining the original image for training with noise using the original image for training and the target image; and training the encoder using the noise image and the original image for training.