Research map / Label-efficient and robust learning / Knowledge distillation
Reinforced Multi-teacher Knowledge Distillation for Efficient General Image Forgery Detection and Localization
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 Re-MTKD, a multi-teacher self-distillation framework with a ConvNeXt-UPerNet student (Cue-Net, with an edge-aware module) whose teacher weights are chosen per batch by a REINFORCE-trained policy, for general image forgery detection and localization.
Problem
Image forgery detection and localization (IFDL) methods generalize poorly across copy-move, splicing and inpainting forgeries, and joint training on mixed tampering types often degrades performance and adds cost.
Approach
Train three forgery-specific Cue-Net teachers and distill them into one student, using a reinforcement-learning policy (logistic policy with REINFORCE) that dynamically selects which teachers to use per batch based on state features from the image, student and teacher outputs, instead of fixed or uniform teacher weights.
Key results
- Average image-level detection AUC of 0.927 and pixel-level localization AUC of 0.861 across copy-move, splicing, inpainting and multi-tampering test sets, the best among compared methods.
- On the multi-tampering set, detection AUC 0.790 vs 0.703 for the next best (HP-FCN), per the main table.
- Ablation: Re-DTS with reward3 and soft3 loss raises average F1 from 0.643 (no teacher) to 0.720 across the detection and localization test sets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Re-MTKD instead of the strongest reported baseline | image forgery detection · Average image forgery detection | AUC: 0.815 -> 0.927 | Helped |
| Use Re-MTKD instead of the strongest reported baseline | image forgery localization · Average image forgery localization | F1: 0.390 -> 0.531 | Helped |
| Add teacher knowledge distillation to Cue-Net (ablation) | Cue-Net · Multi-tampering | Detection F1: 0.764 -> 0.778 | Helped |
| Add teacher knowledge distillation to Cue-Net (ablation) | Cue-Net · Multi-tampering | Localization F1: 0.419 -> 0.444 | Helped |
| Replace fixed equal teacher weights with dynamic teacher selection (ablation) | Cue-Net multi-teacher KD · Average | F1: 0.661 -> 0.720 | Helped |
| Use reward3 instead of reward2 (ablation) | Re-DTS · Average | F1: 0.696 -> 0.720 | Helped |
| Use reward3 instead of reward1 (ablation) | Re-DTS · Average | F1: 0.686 -> 0.720 | Helped |
| Add segmentation soft loss to classification soft loss (ablation) | Re-MTKD with reward3 · Average | F1: 0.660 -> 0.720 | Helped |
| Add PPM to the Cue-Net backbone (ablation) | Cue-Net · Multi-tampering | Average F1: .526 -> .557 | Helped |
| Fuse low-level and high-level EAM features with edge supervision (ablation) | Cue-Net · Multi-tampering | Average F1: .567 -> .591 | Helped |
| Use high-level EAM features instead of low-level EAM features (ablation) | Cue-Net · Multi-tampering | Localization AUC: .863 -> .876 | Helped |
| Use Re-MTKD instead of Cue-Net without KD (ablation) | Cue-Net · Inpainting | Localization F1: 0.530 -> 0.789 | Helped |
Surprising findings
- Uniform-weight multi-teacher ensembling (U-Ensemble) gives a smaller gain and drops on copy-move and multi-tampering detection, suggesting fixed averaging is suboptimal.
- The inpainting-specialized teacher transfers poorly to copy-move and splicing data (copy-move F1 0.000 for Teacher-Inp), indicating a large distribution gap.
Limitations
- Multi-tampering performance remains modest in absolute terms (localization F1 0.444), and the method is comparable to TruFor only after excluding leakage concerns.
- The reported TruFor comparison involves IMD2020 data leakage, and the paper's gains depend on the chosen dataset splits and training data mixture.
Experimental setup
Datasets: CASIA v2 Tampered COCO Fantastic Reality GC Dresden GC Places CASIA v1+ NIST16 Coverage Columbia DSO-1 AutoSplicing DiverseInp
Benchmarks: CASIA v1+ NIST16 Coverage Columbia DSO-1 AutoSplicing DiverseInp IFC Korus IMD2020
Metrics: Accuracy F1 AUC IoU
Baselines: DoaGan Buster-Net CMSD-STRD MFCN RRU-Net CAT-Net HP-FCN IID-Net TLTF-LEFF H-LSTM SPAN MVSS-Net
Builds on: ConvNeXt V2 UPerNet REINFORCE knowledge distillation
Closest papers by meaning
- M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection (AAAI 2025 · similarity 0.87)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.87)
- Stacking Brick by Brick: Aligned Feature Isolation for Incremental Face Forgery Detection (CVPR 2025 · similarity 0.86)
- DiffForensics: Leveraging Diffusion Prior to Image Forgery Detection and Localization (CVPR 2024 · similarity 0.86)
- A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization (AAAI 2025 · similarity 0.86)
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.86)
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.