Scholarlink

Deepfakes created with modern generative models raise serious ethical and forensic concerns. Many detectors still depend on large convolutional networks such as ResNet-50 or VGG-19, which are expensive to run and difficult to deploy on embedded devices. We introduce a compact, factorized CNN, CustomNet-X (CustomNet-X), designed to balance accuracy and efficiency. The network combines: (i) depthwise separable convolutions (DSC) to shrink parameters and FLOPs; (ii) hierarchical residual feedback (HRF) to stabilize gradients and preserve multi-scale cues; and (iii) dynamic feature recalibration (DFR) for channel-wise attention to manipulation-sensitive regions. Trained from scratch on DF Wild Cup, CustomNet-X attains 95.6% accuracy and 0.954 F1 with only 0.42M parameters—about \(\mathbf {8.3\times }\) smaller than MobileNetV2. Generalization holds on FaceForensics++ (91.2%) and Celeb-DF (v2) (92.1%), with robustness to common post-processing. Measured throughput is real time: 312 FPS on an NVIDIA RTX 4070 and 16 FPS on a Jetson Nano. CustomNet-X is therefore a practical, energy-aware choice for edge deployment.

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

CustomNet-X: Factorized CNN Design for Real-Time Deepfake Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Y. Padma Sai,
  • K. Saketh Ram,
  • V. Priyanka Brahmaiah,
  • S. Vaishnavi

摘要

Deepfakes created with modern generative models raise serious ethical and forensic concerns. Many detectors still depend on large convolutional networks such as ResNet-50 or VGG-19, which are expensive to run and difficult to deploy on embedded devices. We introduce a compact, factorized CNN, CustomNet-X (CustomNet-X), designed to balance accuracy and efficiency. The network combines: (i) depthwise separable convolutions (DSC) to shrink parameters and FLOPs; (ii) hierarchical residual feedback (HRF) to stabilize gradients and preserve multi-scale cues; and (iii) dynamic feature recalibration (DFR) for channel-wise attention to manipulation-sensitive regions. Trained from scratch on DF Wild Cup, CustomNet-X attains 95.6% accuracy and 0.954 F1 with only 0.42M parameters—about \(\mathbf {8.3\times }\) smaller than MobileNetV2. Generalization holds on FaceForensics++ (91.2%) and Celeb-DF (v2) (92.1%), with robustness to common post-processing. Measured throughput is real time: 312 FPS on an NVIDIA RTX 4070 and 16 FPS on a Jetson Nano. CustomNet-X is therefore a practical, energy-aware choice for edge deployment.