Skip to content

rediffuse-collaborator-bundle-intake.md

Latest commit

 

History

History
77 lines (54 loc) · 3.99 KB

rediffuse-collaborator-bundle-intake.md

File metadata and controls

77 lines (54 loc) · 3.99 KB

ReDiffuse Collaborator Bundle Intake

Date: 2026-05-09 Status: candidate baseline bundle; not admitted evidence yet

What Arrived

A collaborator provided a DDIMrediffuse folder containing:

  • attack.py
  • components.py
  • model_unet.py
  • dataset_utils.py
  • resnet.py
  • diffusion.py
  • train.py
  • environment.yml
  • CIFAR10_train_ratio0.5.npz
  • CIFAR100_train_ratio0.5.npz

The collaborator later provided train1.py as the training script and stated that the split follows the paper's ratio0.5 npz setting with random seed 42.

The raw bundle is stored outside Git at:

<DIFFAUDIT_ROOT>/Download/shared/supplementary/collaborator-ddim-rediffuse-20260509/

The collaborator-provided 750k checkpoint is stored outside Git at:

<DIFFAUDIT_ROOT>/Download/shared/weights/ddim-cifar10-step750000/

Immediate Findings

  • The earlier single-file attack.py transfer was incomplete; this bundle now supplies the missing modules.
  • components.py defines ReDiffuseAttacker, plus SecMIAttacker, PIA, PIAN, and NaiveAttacker.
  • CIFAR10_train_ratio0.5.npz has SHA256 aca922ecee25ef00dc6b6377ebaf7875dfcc77c2cdfe27c873b26a65134aa0c0, matching the current PIA CIFAR10 ratio0.5 split.
  • model_unet.py can load the collaborator 750k checkpoint ema_model with zero missing keys and zero unexpected keys.
  • A CPU forward probe on the 750k checkpoint produced output shape [1, 3, 32, 32].
  • train1.py uses FLAGS.dataset = "CIFAR10", loads ./CIFAR10_train_ratio0.5.npz, reads mia_train_idxs, and filters torchvision CIFAR10 train data to those indices before training.
  • A Research-side runtime adapter now passes CPU 2/2 and CUDA 8/8 ReDiffuse compatibility smokes. See rediffuse-runtime-smoke-result.md.
  • A CUDA 64/64 small packet has run on the direct-distance scoring surface. See rediffuse-cifar10-small-packet.md.

Interpretation

This bundle is useful. It is not just a loose script; it is a candidate baseline implementation that can align the Research repo with the collaborator's current DDIM/ReDiffuse experiment path.

It is still not admitted evidence because no metric parity has been checked against collaborator-reported results. The current Research packets use a direct first-step distance scorer rather than the collaborator script's second-stage ResNet classifier.

The 750k checkpoint is now better grounded as a CIFAR10 ratio0.5-member-trained target model by collaborator statement plus training-script evidence. Two provenance gaps remain: train1.py does not set seed 42 in code, and its default total_steps = 200000 does not match the checkpoint's stored step = 750000. Treat seed 42 and the 750k training length as collaborator statements until the exact runtime override or final training script is confirmed.

Mainline Decision

Keep the existing 800k PIA CIFAR10 checkpoint as the current gray-box mainline target because its provenance is already workspace-verified through the PIA upstream README-linked pretrained bundle.

Use the 750k checkpoint and DDIMrediffuse bundle for a bounded compatibility review:

  1. Run a minimal ReDiffuse smoke on 750k with a tiny CIFAR10 subset.
  2. If the runner works, map the same protocol onto the existing 800k checkpoint when path layout permits.
  3. Compare ReDiffuse with the existing PIA/SecMI baselines under the same CIFAR10 ratio0.5 split.
  4. Promote only reproduced metrics and low-FPR fields into admitted evidence.

Open Questions

  • Was train1.py run with a command-line or manual override changing total_steps from 200000 to 750000 or 800000?
  • Where was seed 42 set during training?
  • Does train.py reproduce the transferred 750k checkpoint configuration?
  • What metric values does the collaborator expect for ReDiffuse on this checkpoint?

Verdict

Positive intake. The bundle is complete enough to justify a CPU/GPU-light compatibility task, but not enough to replace the existing PIA/SecMI gray-box baseline or to generate a production report.