Research map / Vision-language and multimodal / Multimodal LLMs
SAFE: Semantic- and Frequency-Enhanced Curriculum for Cross-Domain Deepfake Detection
Yulin Yao, Kangfeng Zheng, Bin Wu, Chunhua Wu, Jujie Wang, Jiaqi Gao, Minjiao Yang, Dan Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SAFE, a semantic- and frequency-enhanced curriculum for cross-domain deepfake detection that builds on CLIP's text-image alignment and captions.
Problem
Deepfake detectors trained on one generator family (GANs or diffusion models) degrade substantially when tested on unseen domains, and most CLIP-based detectors use CLIP only as an image-level binary classifier, ignoring its text-to-image semantic alignment.
Approach
Uses CLIP's text-image alignment with captions enriched with high-level semantics (beyond what CLIP-CAP produces), combined with frequency information in a curriculum-style training scheme. Only the abstract was available, so finer details are unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LoRA fine-tuning on CLIP visual branch (ablation) | CLIP ViT-L/14 · CDFv2/DFDCP average | AUC: 0.871 -> 0.941 | Improved |
| add semantic alignment on top of LoRA (ablation) | SAFE · CDFv2/DFDCP average | AUC: 0.941 -> 0.944 | Improved |
| add dual-score curriculum on top of LoRA (ablation) | SAFE · CDFv2/DFDCP average | AUC: 0.941 -> 0.947 | Improved |
| combine alignment and curriculum with LoRA (ablation) | SAFE · CDFv2/DFDCP average | AUC: 0.944 -> 0.956 | Improved |
| augment CLIP-CAP captions with authenticity cues (ablation) | SAFE · 4-set average | AUC: 0.949 -> 0.959 | Improved |
| use raw CLIP-CAP caption instead of fake/real token (ablation) | SAFE · 4-set average | AUC: 0.946 -> 0.949 | Improved |
| set loss weight lambda to 20 instead of 10 or 40 (ablation) | SAFE · CDFv2/DFDCP average | AUC: 0.952 -> 0.956 | Improved |
| use convex pacing vs linear pacing (ablation) | SAFE curriculum · 3 cross-manipulation sets average | AUC: 0.953 -> 0.964 | Improved |
| use convex pacing (ablation) | SAFE curriculum · training time | training time reduction: down 23.3% | Improved |
| initial forged ratio r0=0.3 vs 0.2/0.4/0.5 (ablation) | SAFE · CDFv2/SimSwap average | AUC: 0.958 -> 0.963 | Improved |
| add dual-score curriculum to ResNet-34 | ResNet34 · 5 cross-dataset average (frame-level) | AUC: 0.767 -> 0.807 | Improved |
| add dual-score curriculum to ViT-B/16 | ViT-B16 · 5 cross-dataset average (frame-level) | AUC: 0.845 -> 0.869 | Improved |
Experimental setup
Builds on: CLIP CLIP-CAP
Closest papers by meaning
- Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.90)
- Towards More General Video-based Deepfake Detection through Facial Component Guided Adaptation for Foundation Model (CVPR 2025 · similarity 0.89)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.88)
- Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection (ICML 2025 · similarity 0.88)
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.88)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
