Research map / Image generation / Autoregressive and latent generation
ODDN: Addressing Unpaired Data Challenges in Open-World Deepfake Detection on Online Social Networks
Read the paperCheck your idea against 108,967 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 ODDN, a ResNet-50 deepfake detector that handles mostly unpaired compressed/raw images from OSNs by clustering-based coarse alignment for unpaired data, HSIC for the 20% paired data, and PCGrad-projected adversarial compression discarding, evaluated on 17 GAN datasets.
Problem
Deepfake/AI-generated image detection on online social networks where compressed images are abundant but paired raw originals are scarce, so most training data is unpaired and compression varies across platforms.
Approach
Uses fine-grained HSIC alignment for the small paired subset and coarse cluster-center alignment (real/fake and compressed/uncompressed centers) for unpaired data, plus a gradient-reversal compression-discard branch whose gradients are reconciled with PCGrad so the encoder becomes insensitive to compression.
Key results
- Quality-aware 2-class setting: ODDN reaches 71.4% mean accuracy across 17 datasets vs. 69.2% for QAD, the second best.
- Quality-agnostic 4-class setting: ODDN reaches 72.1% mean accuracy vs. 71.0% for QAD.
- Ablation: baseline 69.4% mean accuracy, +ODA 71.0%, +ODA+CGC 71.4% (quality-aware 2-class).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ODDN instead of QAD | Deepfake detection · 17 GAN datasets, quality-aware, 2-class training | mean accuracy: 69.2 -> 71.4 | Helped |
| Use ODDN instead of QAD | Deepfake detection · 17 GAN datasets, quality-aware, 4-class training | mean accuracy: 70.9 -> 72.6 | Helped |
| Use ODDN instead of QAD | Deepfake detection · 17 GAN datasets, quality-agnostic, 2-class training | mean accuracy: 68.3 -> 70.7 | Helped |
| Use ODDN instead of QAD | Deepfake detection · 17 GAN datasets, quality-agnostic, 4-class training | mean accuracy: 71.0 -> 72.1 | Helped |
| Add ODA (ablation) | Baseline deepfake detector · 17 GAN datasets, quality-aware, 2-class training | mean accuracy: 69.4 -> 71.0 | Helped |
| Add CGC to the baseline with ODA (ablation) | Baseline + ODA · 17 GAN datasets, quality-aware, 2-class training | mean accuracy: 71.0 -> 71.4 | Helped |
Surprising findings
- Baseline methods such as SBIs and MAT drop to near chance on several GANs, and the proposed method's gains over QAD are small (about 2 points mean accuracy).
Limitations
- Gains over the strongest baseline (QAD) are modest, roughly 1.5-2 points mean accuracy.
- Evaluation uses only GAN-generated images with a fixed JPEG quality setting (QP 40) and a sampled quality range, not real OSN-processed images.
- Method description is partly inconsistent (ODA named Isolated Data Aggregation in the method section, DANN referenced in results, and reported 20%/80% split details are ambiguous).
Experimental setup
Datasets: ForenSynths GANGen-Detection
Benchmarks: ForenSynths GANGen-Detection
Metrics: accuracy
Baselines: MesoNet FF++ F3Net MAT SBIs ADD QAD
Builds on: PCGrad HSIC domain-adversarial training ForenSynths
Closest papers by meaning
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.85)
- SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model (CVPR 2025 · similarity 0.85)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.85)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.85)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.85)
- Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.84)
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.