基于动态掩码率的图像篡改定位知识蒸馏算法

黎思力; 廖桂樱; 谭舜泉; 黄继武; 李斌

引用本文：

黎思力,廖桂樱,谭舜泉,黄继武,李斌.基于动态掩码率的图像篡改定位知识蒸馏算法[J].信息安全学报,已采用 [点击复制]
LI Sili,Liao Guiying,Tan Shunquan,Huang Jiwu,Li Bin.Dynamic mask rate-based image tampering localization knowledge distillation algorithm[J].Journal of Cyber Security,Accept [点击复制]

本文已被：浏览 119次下载 0次
基于动态掩码率的图像篡改定位知识蒸馏算法
黎思力, 廖桂樱, 谭舜泉, 黄继武, 李斌
0 字体:加大+\|默认\|缩小-
(深圳大学)

摘要:

尽管基于深度学习的图像篡改定位模型的性能不断提高,但是网络大小和参数量的增加影响了其推理速度和实际部署。知识蒸馏和模型量化是压缩模型的有效方法。其中,MGD(Mask Generation Distillation)是一种使用掩码特征的生成式特征蒸馏算法。然而,在使用该算法对图像篡改定位经典模型DenseFCN进行量化时,不同的掩码率会导致量化模型的精度差异较大。此外,在蒸馏训练中,维持固定的掩码率难以提升模型性能,尤其是在训练后期。因此,本文在MGD算法的基础上提出了一种使用动态掩码率的方法——DMGD(Dynamic MGD)。通过增加一个动态的模块,DMGD算法将掩码率可训练化。此模块将学生特征和教师特征作为输入,在蒸馏过程中根据不同阶段给出适当的逐实例和逐通道的掩码率。实验结果表明,在Defacto、PS-boundary等数据集上,相较于MGD和未使用蒸馏方式,DMGD能够使DenseFCN低比特量化模型在F1等指标上达到最优。

关键词: 深度学习图像篡改定位模型量化知识蒸馏

DOI：

投稿时间：2023-12-12修订日期：2024-06-11

基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目）

Dynamic mask rate-based image tampering localization knowledge distillation algorithm

LI Sili, Liao Guiying, Tan Shunquan, Huang Jiwu, Li Bin

(Shenzhen University)

Abstract:

In recent years, the size and number of parameters in deep learning-based models for image tampering localization have been progressively increasing. Compressing these models has been effective using methods such as Knowledge distillation and model quantization. One of these algorithms, Mask Generation Distillation (MGD), is a generative feature distillation that make use of masking features. However, when applying MGD to classic image tampering localization model DenseFCN, the accuracy of the quantized model can be significantly affected by varying masking rates. Additionally, it becomes difficult to enhance the model’s performance in the later stages of traing when a fixed masking rate is maintained during distillation. Therefore, in this paper, we introduce Dynamic MGD (DMGD), a method which builds upon the MGD algorithm and employs a dynamic masking rate. The DMGD algorithm includes a dynamic module that makes the masking rate trainable. This module takes the student and teacher features as inputs and provides instance-wise and channel-wise masking rates which are appropriate for different stages during the distillation process. Results on public datasets such as Defacto and PS-boundary, demonstrate that DMGD achieves the optimal performance on metrics such as F1 score, compared to MGD and non-distillation methods, for low-bit quantized DenseFCN models.

Key words: deep learning, image tampering localization, modelquantization, knowledge distillation