Image encryption method for target detection task

By loading and freezing the target detection model to generate encrypted images, and optimizing the encrypted images using a joint loss function and momentum superposition strategy, the problem of balancing image privacy protection and performance in target detection tasks is solved. This achieves an efficient image encryption method that maintains detection accuracy and is easy to integrate.

CN121010847APending Publication Date: 2025-11-25SHAANXI NORMAL UNIV
0 Cites 0 Cited by

Patent Information

Application Number
CN202511056733.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-30
Publication Date
2025-11-25

AI Technical Summary

Technical Problem

Existing technologies lack efficient image encryption methods that can balance image privacy protection and target detection performance, especially in cloud or edge deployments where it is difficult to maintain detection performance while protecting image privacy.

Method used

An initial encrypted image is generated by loading an object detection model and freezing its parameters. The pixel values ​​of the encrypted image are then optimized using a joint loss function and multiple rounds of training. The encrypted image is updated to maintain detection consistency and feature consistency by combining momentum stacking and learning rate adaptive strategies.

Benefits of technology

It enables the target detection model to correctly parse even visually unrecognizable encrypted images, maintaining detection performance with a slight decrease in detection accuracy, and is easy to integrate into existing systems.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention relates to an image encryption method for a target detection task, which comprises the following steps of: loading an original image from a data set and preprocessing the original image, including data enhancement, normalization and size adjustment; loading the target detection model D, and freezing all parameters of the target detection model D; generating a corresponding initial encrypted image based on the loaded original image and setting the initial encrypted image in a training state; in the first round of training, the original image and the initial encrypted image serve as input of a frozen target detection model D, and the target detection model D respectively extracts interlayer features of the two images and outputs a detection result at the same time; based on the extracted middle layer features and the detection result, back propagation and gradient calculation are carried out on the encrypted image by using a joint loss function lambda, and the pixel value of the encrypted image is updated according to the gradient to obtain an optimized encrypted image; and from the second round, each round of training takes the original image and the encrypted image optimized in the previous round as input, and a final encrypted image is obtained through multiple rounds of training until a specific stop condition is met.
Need to check novelty before this filing date? Find Prior Art