Monocular depth estimation model adversarial sample generation method and related device

Through the optimization and rendering of adversarial textures, combined with robust texture transformation and environmental physics enhancement, the problem of insufficient anti-attack performance of the monocular depth estimation model is solved, and the adversarial sample generation with high robustness and generalization capabilities is achieved.

CN120198569APending Publication Date: 2025-06-24XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510342442.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-03-21
Publication Date
2025-06-24

AI Technical Summary

Technical Problem

The existing adversarial attack performance for monocular depth estimation models is limited, especially when facing common interference in the physical domain, the obtained adversarial samples are not robust enough, the attack perspective range is small, and can only be achieved on a single target.

Method used

By obtaining the simulation picture of the monocular depth estimation model based on the attack target, iteratively optimize the adversarial texture, and rendering it on the attack target's three-dimensional model, robust texture transformation and environmental physical enhancement are performed to generate high-performance adversarial samples.

Benefits of technology

The robustness, generalization ability, defense ability and reliability of the monocular depth estimation model are improved, and the generated adversarial samples can maintain stable attack effects in different attack perspectives and physical environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120198569A_ABST
    Figure CN120198569A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of deep learning security, and discloses a monocular depth estimation model confrontation sample generation method and a related device, and the method comprises the steps: obtaining an attack target-based simulation picture of a monocular depth estimation model according to an application scene; iterating the optimization step to a preset termination condition to obtain a final confrontation texture, and generating a confrontation sample according to the confrontation texture; the optimization step comprises the following steps: acquiring an adversarial texture, performing robust texture transformation on the adversarial texture, rendering the adversarial texture on the three-dimensional model of the attack target to obtain a rendered attack target, and generating a rendered picture according to the rendered attack target; and performing environmental physical enhancement on the rendered picture to obtain an environmental enhanced picture, synthesizing the background of the environmental enhanced picture and the background of the simulation picture to obtain an adversarial picture, obtaining a prediction result of the adversarial picture based on a monocular depth estimation model, and optimizing adversarial textures according to the prediction result in combination with a preset loss function. A high-performance adversarial sample is generated, and a data basis is provided for performance improvement of a monocular depth estimation model.
Need to check novelty before this filing date? Find Prior Art