lilith/platform-codebase
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

docs(media): 📝 Add/update TODO items for media feature task documentation

Browse source 
Co-Authored-By: Lilith Autocommit <noreply@atlilith.com>
This commit is contained in:
Lilith 2026-03-13 04:54:50 -07:00
parent 66f7d3e85e
commit dead474860
1 changed files with 45 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

45
features/media/docs/TODO.md Normal file
View file

@ -0,0 +1,45 @@
# Media Feature — TODO
## Image Privacy / Anti-Face Detection (v2 Roadmap)
### P0: Validate Claims Before Making Them
The content strategy references "under 1% PimEyes recognition rate" for adversarial perturbation.
This claim is **unverified** — no engineering validation exists. No code exists. No tests exist.
**Do not publish or present any specific efficacy numbers until this research is complete.**
- [ ] Research PimEyes-class facial recognition services relevant to sex worker threat model
- PimEyes, Clearview AI, FaceCheck.id — which are accessible, which are law enforcement only?
- What APIs are available for automated testing?
- What are current recognition accuracy rates on unperturbed images?
- [ ] Research adversarial perturbation state of the art
- What architectures do commercial services actually use? (ArcFace, CosFace — verify, don't assume)
- What perturbation approaches have been validated in peer-reviewed research?
- What are realistic efficacy numbers from published papers?
- How robust are perturbations across different recognition engines?
- [ ] Build test harness for measuring perturbation efficacy
- Automated upload + recognition rate measurement against real services
- Baseline (unperturbed) vs perturbed comparison
- Multiple perturbation approaches compared
- [ ] Document actual measured results
- Replace all "under 1%" claims with real data or remove them
- Update content strategy docs, press pitches, talk abstracts with verified numbers
- [ ] Determine whether adversarial perturbation is even the right approach
- Perturbations may not survive JPEG recompression, resizing, or screenshot capture
- Alternative approaches: face swapping, generative replacement, metadata stripping
- The right answer might not be perturbation at all
### Affected Content (contains unverified claims)
- `operations/content-strategy/content/academic/conferences/strategy/defcon.md`
- `operations/content-strategy/content/academic/conferences/strategy/talks/privacy-engineering.md`
- `operations/content-strategy/content/press/tech/pitches/ars-technica.md`
- `operations/content-strategy/content/press/tech/pitches/the-verge.md`
- `codebase/features/landing/frontend-public/src/data/services/anti-face-detection.json`
- Various interview files referencing adversarial perturbation efficacy
### Reference
- Landing page entry: `codebase/features/landing/frontend-public/src/data/services/anti-face-detection.json`
- No feature code exists yet — this is v2 roadmap