Deep learning training picture encryption and decryption method

Abstract

The invention relates to a deep learning training picture encryption and decryption method. The method comprises an encryption process and a decryption process, wherein the encryption process comprises the steps of converting read picture data into a three-dimensional array, encrypting three channels of each picture, exchanging the encrypted three channels, and then combining the three channels toobtain an output picture. The decryption process comprises the following steps of: acquiring pictures to be decrypted, performing channel exchange on each picture, decrypting the three exchanged channels respectively, and combining the three channels to obtain an output picture. According to the method, model leakage and picture data leakage can be prevented, and the target detection model can betrained.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method for encrypting and decrypting training pictures for deep learning. Background technique [0002] The existing conventional solution is plaintext image training. The current image encryption technology can be divided into two categories, namely airspace image encryption technology and compressed image encryption technology. The spatial domain image encryption technology encrypts the uncompressed image, which is characterized by treating the image as three-dimensional data (length and width plus channel) for operation. The typical method of airspace image encryption technology is to use discrete chaos encryption technology. Compressed image encryption technology is encrypted based on a certain compression format or compression technology, such as JPEG, arithmetic coding, wavelet compression technology, etc. [0003] These encryption techniques will lead to non-...

AI Technical Summary

Problems solved by technology

This scheme can only obtain the dimensionality reduction vector features of the picture by encrypting the target picture, while the neural network of the target detection requires the regression of the target position and the classification of the target on the picture, and the local encryption scheme loses the data of the target to be detected. The spatial position information on a picture, and the data that is dimensionally reduced at the same time cannot be spliced ​​back to the original picture

Therefore, traditional feature encryption or partial encryption cannot meet the dual requirements of target detection and target classification.

It can be seen that the data encrypted by local features has no spatial features of the image target position, and cannot be trained

[0005] The core objective of the present invention is to solve the problem of data leakage caused by power safety operation data being given to third-party manufacturers for training and the problem of inability to train due to the loss of space characteristics of encrypted data

Images