Hi @xuanyuzhang21,
Congratulations on EditGuard's CVPR 2024 acceptance! The tamper localization results look impressive.
SiliconSignature — Hardware-Bound Watermarking
I'd like to propose a complementary approach that addresses a different threat model.
How We Differ (And Complement)
| Aspect |
EditGuard |
SiliconSignature |
| Forge cost |
Retrain model |
$10,000+ (buy ASIC) |
| Hardware bound |
No |
Yes (Antminer S9) |
| Tamper detection |
Localization mask |
Exact pixel changes (RS syndrome) |
| Offline verify |
Requires model |
Math only (no ML) |
| Target |
AIGC editing |
Any image (camera, AI, scan) |
Proposed Collaboration
Two-layer defense:
- EditGuard: Proactive watermarking for AIGC-generated images (your focus)
- SiliconSignature: Hardware-bound signing for any image source (our focus)
When an image has both watermarks:
- EditGuard proves it was proactively protected
- SiliconSignature proves the specific ASIC that signed it
- Tampering breaks both = maximum confidence
Specific Integration Idea
Could we add a SiliconSignature verification node to your inference pipeline? After EditGuard detects tampering, SiliconSignature could:
- Verify if the image was ever signed
- Reveal which pixels were modified (via Reed-Solomon syndrome)
- Provide hardware identity of the original signer
Technical Details
Author: Francisco Angulo de Lafuente (@Agnuxo1) — 35 years research
— Francisco
Hi @xuanyuzhang21,
Congratulations on EditGuard's CVPR 2024 acceptance! The tamper localization results look impressive.
SiliconSignature — Hardware-Bound Watermarking
I'd like to propose a complementary approach that addresses a different threat model.
How We Differ (And Complement)
Proposed Collaboration
Two-layer defense:
When an image has both watermarks:
Specific Integration Idea
Could we add a SiliconSignature verification node to your inference pipeline? After EditGuard detects tampering, SiliconSignature could:
Technical Details
Author: Francisco Angulo de Lafuente (@Agnuxo1) — 35 years research
— Francisco