Research map / Image generation / Image restoration
FiLM-CAM: Keyed Feature Modulation for Conditional-Access Watermarking
Yash Kulthe, Vishal Asnani, Shruti Agarwal, Andrew Gilbert, John Collomosse
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 FiLM-CAM, a conditional-access image watermarking method in which a secret key drives feature-wise linear modulation of the watermark signal's internal representation, so that encoding, decoding, and removal all require the matching key.
Problem
Image watermarking typically lets anyone with the detector read or remove a watermark; the paper targets conditional access, where encoding, decoding and removal should each be restricted to authorized key holders.
Approach
A secret key conditions a FiLM mechanism that modulates the internal feature representation of the watermark before embedding, coupling the payload to the key at the feature level rather than the pixel level.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FiLM conditioning with direct key input to the embedder (ablation) | FiLM-CAM · DIV2K | bit accuracy: 0.85 -> 0.99 | Improved |
| Add FiLM conditioning (ablation) | FiLM-CAM · DIV2K | removal PSNR (dB): 47.3 -> 55.4 | Improved |
| Add the image-dependent gain term (ablation) | FiLM-CAM · DIV2K | removal PSNR (dB): 44.0 -> 55.4 | Improved |
| Add the gain term (ablation) | FiLM-CAM · DIV2K | bit accuracy: 0.99 -> 0.99 | No clear effect |
| Use logit-space composition instead of additive composition (ablation) | FiLM-CAM · DIV2K | removal PSNR (dB): 54.1 -> 55.4 | Improved |
| Use logit-space composition instead of additive composition (ablation) | FiLM-CAM · COCO | removal PSNR (dB): 55.5 -> 59.3 | Improved |
| Use logit-space composition instead of additive composition (ablation) | FiLM-CAM · DIV2K | watermarked PSNR (dB): 47.0 -> 46.7 | No clear effect |
| Use ConvNeXt-Base extractor instead of ResNet-50 (ablation) | FiLM-CAM encoder · DIV2K | watermarked PSNR (dB): 37.5 -> 46.7 | Improved |
| Add noise layer during training (ablation) | FiLM-CAM · COCO | watermarked PSNR (dB): 48.2 -> 40.2 | Hurt |
| Add noise layer during training (ablation) | FiLM-CAM · Wrong-key removal | post-removal bit accuracy: 0.81 wrong-key survival | Improved |
| Increase iterative refinement steps beyond 20 for removal (ablation) | FiLM-CAM · removal PSNR | removal PSNR | No clear effect |
| Repeatedly re-embed watermarks without removal (ablation) | FiLM-CAM · Repeated Watermark Embedding | PSNR and bit accuracy | Hurt |
Experimental setup
Builds on: FiLM
Closest papers by meaning
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.84)
- END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions (AAAI 2025 · similarity 0.84)
- RoPaSS: Robust Watermarking for Partial Screen-Shooting Scenarios (AAAI 2025 · similarity 0.83)
- Decoder Gradient Shield: Provable and High-Fidelity Prevention of Gradient-Based Box-Free Watermark Removal (CVPR 2025 · similarity 0.83)
- Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft (CVPR 2025 · similarity 0.83)
- RAW: A Robust and Agile Plug-and-Play Watermark Framework for AI-Generated Images with Provable Guarantees (NeurIPS 2024 · similarity 0.83)
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.
