Method for improving robustness of image compression algorithm based on adversarial attack

Abstract

The invention discloses a method for improving robustness of an image compression algorithm based on adversarial attacks. The method comprises the following steps: obtaining a to-be-tested neural network image compression algorithm model and parameters; for any input picture, adding the initial disturbance into the input picture to obtain a reconstructed picture; iteratively optimizing disturbance by taking maximization of output loss or minimization of output of a gap between the reconstructed picture and the target picture as a target, and generating an adversarial sample; and integrating the confrontation samples into training of a network model until the model is fully robust to disturbance. The method provided by the invention is a method for improving robustness of various neural network compression algorithms, and in equivalent comparison with a traditional neural network-based image compression method on a large number of image sequences, the PSNR (Peak Signal to Noise Ratio) performance of more than 10dB can be improved in the face of adversarial attacks.

Description

technical field [0001] The invention relates to the field of image coding, in particular to a method for improving the robustness of an image compression algorithm based on adversarial attacks. Background technique [0002] In recent years, artificial neural networks have developed to the stage of deep learning. Deep learning attempts to use a series of algorithms that contain complex structures or multiple processing layers composed of multiple nonlinear transformations to perform high-level abstraction on data. Its powerful expressive ability enables it to achieve the best results in various machine learning tasks. The performance on video and image processing also currently exceeds other methods. [0003] Deep learning uses the idea of ​​hierarchical abstraction, and high-level concepts are learned through low-level concepts. This hierarchical structure is usually constructed using a greedy layer-by-layer training algorithm, and effective features that are helpful for m...

AI Technical Summary

Problems solved by technology

The current mainstream image compression methods based on deep learning are all based on convolutional neural network. Due to the inherent characteristics of convolutional neural network itself, the transformation used as an autoencoder has certain instability, making it possible to resist attacks.

Images