Initialize DEFUSE checkpoint release

README.md

@@ -0,0 +1,60 @@
+---
+tags:
+- computer-vision
+- diffusion
+- sdxl
+- clip
+- backdoor-detection
+- research
+base_model:
+- stabilityai/stable-diffusion-xl-base-1.0
+- openai/clip-vit-base-patch32
+---
+
+# DEFUSE checkpoints
+
+Official checkpoints for **DEFUSE: Generalizable Backdoor Defense for Self-Supervised Encoders with Generative Priors** (ACM Multimedia 2026, [paper](https://doi.org/10.1145/3767308.3835471)).
+
+These are custom PyTorch checkpoints used by the DEFUSE codebase; they are not a standalone Diffusers pipeline.
+
+## Files
+
+| Path | Description | Size | SHA-256 |
+|---|---|---:|---|
+| `clip-backdoor/epoch.best.pt` | CLIP ViT-B/32 backdoored encoder checkpoint used by the final experiment (`state_dict`, epoch 6) | 1,815,827,343 bytes | `0820c39e771499158c67084681955407ae8ab931c03cc0f520c8b670daf3a14e` |
+| `sdxl-defuse/final.pt` | Final DEFUSE SDXL conditioning weights at step 30,000; this is not the full SDXL base model | 1,790,009,754 bytes | `172b1b5bb20488dea04c33a7216c51c9fa0050fbc8bfe007f19759357e4ee17b` |
+| `training/config.resolved.yaml` | Fully resolved configuration for the final experiment | - | - |
+| `training/run.json` | Runtime and environment metadata | - | - |
+
+The final experiment was `in900-best-encoder-no-grad-clip-30k-20260818`, trained on ImageNet-900 for 30,000 optimizer steps. The SDXL checkpoint metadata is:
+
+```text
+model_type: sdxl
+image_tokens: 4
+feature_dim: 512
+feature_mode: global
+cross_attention_dim: 2048
+include_text_tokens: false
+conditioning_version: 2
+```
+
+## Use with DEFUSE
+
+Download the repository, then point the DEFUSE configuration to:
+
+```yaml
+encoder:
+  model: openai/clip-vit-base-patch32
+  checkpoint: /path/to/clip-backdoor/epoch.best.pt
+
+reconstruction:
+  checkpoint: /path/to/sdxl-defuse/final.pt
+```
+
+You must also provide the SDXL base model configured by the DEFUSE project. See the project README and training configuration for the complete environment and data settings.
+
+## Responsible use and licenses
+
+The CLIP checkpoint is intentionally backdoored and is released only for defensive security research and reproducibility. Do not deploy it as a trusted production encoder.
+
+The checkpoints derive from third-party base models and training code. Users are responsible for complying with the applicable upstream licenses and terms, including those of SDXL, OpenAI CLIP, and the source backdoor-training implementation.

training/config.resolved.yaml

@@ -0,0 +1,72 @@
+seed: 42
+model:
+  pretrained: /workspace/pretrained_models/sdxl
+  type: sdxl
+  image_tokens: 4
+  include_text_tokens: false
+encoder:
+  model: /workspace/pretrained_models/clip-vit-base-patch32
+  type: hf_clip
+  checkpoint: /workspace/SSL-Backdoor/ssl_backdoor/projects/CleanCLIP/logs/clip_10epoch_finetune/checkpoints/epoch.best.pt
+  feature_mode: global
+data:
+  train_manifest: /tmp/defuse_in900_train.jsonl
+  val_manifest: /tmp/defuse_in900_val.jsonl
+  root: /workspace/dataset/imagenet-1k
+  resolution: 224
+  center_crop: true
+  num_workers: 0
+training:
+  output_dir: /tmp/defuse_in900_best_encoder_30k_20260818
+  batch_size: 1
+  learning_rate: 0.0001
+  weight_decay: 0.01
+  max_steps: 30000
+  gradient_accumulation_steps: 11
+  mixed_precision: fp16
+  checkpoint_interval: 3000
+  checkpoint_keep_last: 2
+  eval_interval: 2000
+  max_grad_norm: null
+  resume: null
+evaluation:
+  dino_model: /workspace/pretrained_models/dinov2-base
+  num_samples: 100
+  batch_size: 1
+  run_at_start: true
+  run_at_end: true
+  seed: 1234
+  inference_steps: 25
+  save_images: 4
+  dino_threshold: 0.1
+reconstruction:
+  checkpoint: null
+  manifest: /tmp/defuse_in900_val.jsonl
+  output_dir: /tmp/defuse_in900_best_encoder_30k_20260818-reconstruct
+logging:
+  backends:
+  - local
+  - wandb
+  level: INFO
+  train_interval: 1
+  train_dino_interval: 200
+  wandb:
+    project: defuse
+    entity: jsrdcht
+    name: in900-best-encoder-no-grad-clip-30k-20260818
+    group: in900-epoch1-regression
+    tags:
+    - formal-test
+    - refactored
+    - dual-attention
+    - cached-empty-text
+    - sdxl
+    - clip-backdoor
+    - in900
+    - train-dino
+    - best-encoder
+    - no-grad-clip
+    - 30k
+    mode: online
+    run_id: null
+    resume: auto

training/run.json

@@ -0,0 +1,36 @@
+{
+  "timestamp": "2026-08-18T08:10:33Z",
+  "command": [
+    "/workspace/DEFUSE/src/defuse/train.py",
+    "--config",
+    "/workspace/DEFUSE/tmp/formal_best_encoder_30k_20260818/config.yaml"
+  ],
+  "git": {
+    "status": "not-versioned"
+  },
+  "python": "3.10.20",
+  "platform": "Linux-5.15.0-139-generic-x86_64-with-glibc2.35",
+  "world_size": 6,
+  "device": {
+    "cuda_available": true,
+    "cuda_device_count": 6,
+    "cuda_devices": [
+      "NVIDIA GeForce RTX 4090",
+      "NVIDIA GeForce RTX 4090",
+      "NVIDIA GeForce RTX 4090",
+      "NVIDIA GeForce RTX 4090",
+      "NVIDIA GeForce RTX 4090",
+      "NVIDIA GeForce RTX 4090"
+    ],
+    "cuda_visible_devices": "0,1,2,3,4,5"
+  },
+  "versions": {
+    "accelerate": "1.13.0",
+    "diffusers": "0.36.0",
+    "omegaconf": "2.3.0",
+    "torch": "2.4.1+cu118",
+    "torchvision": "0.19.1+cu118",
+    "transformers": "4.57.6",
+    "wandb": "0.26.1"
+  }
+}