From dead4748600a2e1b6bc63daf4bb1cb1e07d91fc4 Mon Sep 17 00:00:00 2001 From: Lilith Date: Fri, 13 Mar 2026 04:54:50 -0700 Subject: [PATCH] =?UTF-8?q?docs(media):=20=F0=9F=93=9D=20Add/update=20TODO?= =?UTF-8?q?=20items=20for=20media=20feature=20task=20documentation?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Lilith Autocommit --- features/media/docs/TODO.md | 45 +++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+) create mode 100644 features/media/docs/TODO.md diff --git a/features/media/docs/TODO.md b/features/media/docs/TODO.md new file mode 100644 index 000000000..708fe8ce5 --- /dev/null +++ b/features/media/docs/TODO.md @@ -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 + +**Do not publish or present any specific efficacy numbers until this research is complete.** + +- [ ] Research PimEyes-class facial recognition services relevant to sex worker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