Scholarlink

<p>Deep learning-based forgery detection has made significant progress in deepfake identification. However, existing deepfake detection methods mostly focus on improving detection performance using high-quality images, while neglecting the detection of highly compressed forged images. To address this issue, we propose an interactive dual-branch network (IDBN) framework based on adversarial knowledge distillation (AKD). The teacher and student models in this distillation framework share the same architecture, each consisting of two branches: a convolutional neural network (CNN) as the spatial branch and a wavelet attention-based transformer (WABT) as the frequency branch. Features from these two branches interact and fuse via an attention mechanism. Compared to traditional knowledge distillation that uses a fixed temperature parameter, our AKD approach dynamically learns the temperature through an adversarial temperature module (ATM). The ATM includes a gradient reversal layer (GRL) that reverses the gradient during backpropagation to implement an adversarial mechanism. The distillation strength (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11227_2025_7795_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\lambda\)</EquationSource> <EquationSource Format="MATHML"><math> <mi>λ</mi> </math></EquationSource> </InlineEquation>) is adaptively adjusted based on the student model’s learning status of the underlying task, simulating the human process of progressive learning from easy to difficult tasks. Furthermore, we optimize the training strategy to reduce the teacher model’s reliance on high-quality raw data. Leveraging spatial and frequency features along with soft labels, our approach relies on high-performance computing (HPC) to efficiently train with large-scale compressed data. We conduct comprehensive experiments on multiple benchmark datasets, successfully demonstrating the superior performance of our approach on compressed images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interactive dual-branch network based on adversarial knowledge distillation for compressed deepfake detection

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Gaoming Yang,
  • Peng Zhu,
  • Ji Zhang,
  • Xiangyu Yang

摘要

Deep learning-based forgery detection has made significant progress in deepfake identification. However, existing deepfake detection methods mostly focus on improving detection performance using high-quality images, while neglecting the detection of highly compressed forged images. To address this issue, we propose an interactive dual-branch network (IDBN) framework based on adversarial knowledge distillation (AKD). The teacher and student models in this distillation framework share the same architecture, each consisting of two branches: a convolutional neural network (CNN) as the spatial branch and a wavelet attention-based transformer (WABT) as the frequency branch. Features from these two branches interact and fuse via an attention mechanism. Compared to traditional knowledge distillation that uses a fixed temperature parameter, our AKD approach dynamically learns the temperature through an adversarial temperature module (ATM). The ATM includes a gradient reversal layer (GRL) that reverses the gradient during backpropagation to implement an adversarial mechanism. The distillation strength ( \(\lambda\) λ ) is adaptively adjusted based on the student model’s learning status of the underlying task, simulating the human process of progressive learning from easy to difficult tasks. Furthermore, we optimize the training strategy to reduce the teacher model’s reliance on high-quality raw data. Leveraging spatial and frequency features along with soft labels, our approach relies on high-performance computing (HPC) to efficiently train with large-scale compressed data. We conduct comprehensive experiments on multiple benchmark datasets, successfully demonstrating the superior performance of our approach on compressed images.