Upload folder using huggingface_hub

.gitattributes

@@ -35,3 +35,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 agent_fp16.trt filter=lfs diff=lfs merge=lfs -text
 agent_fp32.trt filter=lfs diff=lfs merge=lfs -text
+paper.pdf filter=lfs diff=lfs merge=lfs -text
+tensorrt/rfpar_cls_v1_fp16.trt filter=lfs diff=lfs merge=lfs -text
+tensorrt/rfpar_cls_v1_fp32.trt filter=lfs diff=lfs merge=lfs -text
+tensorrt/rfpar_det_v1_fp16.trt filter=lfs diff=lfs merge=lfs -text
+tensorrt/rfpar_det_v1_fp32.trt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,62 +1,107 @@
 ---
-library_name: pytorch
 tags:
+  - robotics
+  - anima
   - adversarial-attacks
   - reinforcement-learning
   - pixel-attack
   - image-classification
   - object-detection
-  - anima
-license: mit
+  - robot-flow-labs
+library_name: pytorch
+pipeline_tag: robotics
+license: apache-2.0
 datasets:
   - imagenet-1k
   - coco
 ---
 
-# DEF-rfpar: RFPAR — Remember and Forget Pixel Attack using RL
+# DEF-rfpar — RFPAR: Remember and Forget Pixel Attack using RL
+
+Part of the [ANIMA Perception Suite](https://robotflowlabs.com) by Robot Flow Labs.
 
-ANIMA defense module implementing the RFPAR adversarial attack algorithm from [ArXiv 2502.07821](https://arxiv.org/abs/2502.07821).
+## Paper
+
+**RFPAR: Remember and Forget Pixel Attack using Reinforcement Learning** ([ArXiv 2502.07821](https://arxiv.org/abs/2502.07821))
+Dongsu Song, Daehwa Ko, Jay Hoon Jung — Korea Aerospace University
+
+## Architecture
 
-## Model Description
+RFPAR uses a REINFORCE policy network (Conv2d + FC) to select optimal pixel perturbations for black-box adversarial attacks. The key innovation is the **Remember and Forget** process:
 
-RFPAR uses a REINFORCE policy network to select optimal pixel perturbations for black-box adversarial attacks. The "Remember and Forget" process:
-- **Remember**: Train RL agent to find effective pixel perturbations
-- **Forget**: Reset agent, keep successful perturbations in memory
+- **Remember**: Train RL agent to find effective pixel perturbations via one-step policy gradient
+- **Forget**: Reset agent parameters, keep successful perturbations in memory, re-optimize from new starting point
+
+Two variants are provided:
+- **Classification** (`rfpar_cls_v1`): Attacks image classifiers (ResNeXt50 on ImageNet)
+- **Detection** (`rfpar_det_v1`): Attacks object detectors (YOLO on COCO)
 
 ## Results
 
 ### Classification (ImageNet-1K, ResNeXt50)
+
 | Metric | Value |
 |--------|-------|
-| Attack Success Rate | 93.0% |
-| Images Deceived | 186/200 |
+| Attack Success Rate | **93.0%** |
+| Images Deceived | 186 / 200 |
 | Mean L0 | 143.9 |
 | Mean L2 | 6.26 |
-| Avg Queries | 463 |
+| Average Queries | 463 |
 | Forget Iterations | 100 |
-| Time | 419s (NVIDIA L4) |
 
-## Available Formats
+### Detection (COCO, YOLO11n)
 
-- `agent.pth` — PyTorch checkpoint
-- `agent.safetensors` — Safetensors format
-- `agent.onnx` — ONNX format (opset 18)
-- `agent_fp16.trt` — TensorRT FP16
-- `agent_fp32.trt` — TensorRT FP32
+| Metric | Value |
+|--------|-------|
+| Boxes Eliminated | 169 total |
+| Box Elimination Rate | **112%** |
+| Mean L0 | 234.8 |
+| Mean L2 | 1885.2 |
+| Average Queries | 494 |
+| Forget Iterations | 5 |
+
+## Exported Formats
+
+| Format | Classification | Detection | Use Case |
+|--------|---------------|-----------|----------|
+| PyTorch (.pth) | `pytorch/rfpar_cls_v1.pth` | `pytorch/rfpar_det_v1.pth` | Training, fine-tuning |
+| SafeTensors | `pytorch/rfpar_cls_v1.safetensors` | `pytorch/rfpar_det_v1.safetensors` | Fast loading, safe |
+| ONNX | `onnx/rfpar_cls_v1.onnx` | `onnx/rfpar_det_v1.onnx` | Cross-platform inference |
+| TensorRT FP16 | `tensorrt/rfpar_cls_v1_fp16.trt` | `tensorrt/rfpar_det_v1_fp16.trt` | Edge deployment (Jetson/L4) |
+| TensorRT FP32 | `tensorrt/rfpar_cls_v1_fp32.trt` | `tensorrt/rfpar_det_v1_fp32.trt` | Full precision inference |
 
 ## Usage
 
 ```python
-from anima_rfpar.export import load_agent_from_checkpoint
+import torch
+from anima_rfpar.agent import REINFORCEAgent
+
+# Load classification agent
+agent = REINFORCEAgent(224, 224, 3, detector_mode=False)
+ckpt = torch.load("pytorch/rfpar_cls_v1.pth", weights_only=False)
+agent.load_state_dict(ckpt["agent_state_dict"])
+agent.eval()
 
-agent = load_agent_from_checkpoint("best.pth")
-mean, std = agent(image_tensor)  # Returns (x, y, r, g, b) action distribution
+# Get pixel perturbation action for an image
+image = torch.randn(1, 3, 224, 224)  # [0, 1] normalized
+action_mean, action_std = agent(image)
+# action_mean: (1, 5) -> sigmoid -> (x, y, r, g, b)
 ```
 
-## Paper
+## Training
+
+- **Hardware**: NVIDIA L4 (23GB VRAM)
+- **CUDA**: 12.0 / PyTorch cu128
+- **Framework**: PyTorch 2.11
+- **Config**: See `configs/` directory
+- **Time**: 419s (classification), 1581s (detection)
 
-Song, D., Ko, D., & Jung, J. H. (2025). RFPAR: Remember and Forget Pixel Attack using Reinforcement Learning. *arXiv:2502.07821*.
+## Product Stack
+
+ATLAS / ORACLE — Defense Marketplace
 
 ## License
 
-MIT — Built with ANIMA by Robot Flow Labs
+Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED
+
+Built with ANIMA by Robot Flow Labs

checkpoints/best_cls.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b99506f7ce6615a97d11865a8a2c79681ab50899fd70bdfc2784b724f938b0c
+size 411145941

checkpoints/best_det.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8cd89dd3eeb73931cd47f6fc2ff7235556da97f32ed8e7d3281b42c409e63d5
+size 411145941

configs/paper_cls.toml

@@ -0,0 +1,23 @@
+[module]
+name = "DEF-rfpar"
+seed = 2
+
+[execution]
+backend = "cuda"
+mode = "classification"
+dry_run = false
+output_dir = "/mnt/artifacts-datai/checkpoints/DEF-rfpar"
+
+[paths]
+reference_repo = "repositories/RFPAR"
+imagenet_root = "repositories/RFPAR/ImageNet"
+coco_root = "repositories/RFPAR/COCO/images/val"
+argoverse_root = "/mnt/forge-data/datasets/argoverse"
+
+[attack]
+max_iterations = 100
+bound_threshold_eta = 0.05
+alpha = 0.01
+convergence_duration_t = 3
+query_budget = 1700
+yolo_conf_threshold = 0.5

configs/paper_det.toml

@@ -0,0 +1,23 @@
+[module]
+name = "DEF-rfpar"
+seed = 2
+
+[execution]
+backend = "cuda"
+mode = "detection"
+dry_run = false
+output_dir = "/mnt/artifacts-datai/checkpoints/DEF-rfpar"
+
+[paths]
+reference_repo = "repositories/RFPAR"
+imagenet_root = "repositories/RFPAR/ImageNet"
+coco_root = "repositories/RFPAR/COCO/images/val"
+argoverse_root = "/mnt/forge-data/datasets/argoverse"
+
+[attack]
+max_iterations = 100
+bound_threshold_eta = 0.05
+alpha = 0.05
+convergence_duration_t = 20
+query_budget = 1700
+yolo_conf_threshold = 0.5

logs/metrics_cls.json

@@ -0,0 +1,11 @@
+{
+  "mode": "classification",
+  "total_images": 200,
+  "total_deceived": 186,
+  "success_rate": 0.93,
+  "mean_l0": 143.86559139784947,
+  "mean_l2": 6.261905570824941,
+  "mean_queries": 462.595,
+  "forget_iterations": 100,
+  "elapsed_sec": 419.3310854434967
+}

logs/metrics_det.json

@@ -0,0 +1,11 @@
+{
+  "mode": "detection",
+  "total_images": 50,
+  "total_deceived": 169,
+  "success_rate": 1.1192053359075886,
+  "mean_l0": 234.76,
+  "mean_l2": 1885.1656958007814,
+  "mean_queries": 494.0,
+  "forget_iterations": 5,
+  "elapsed_sec": 1581.1591651439667
+}

onnx/rfpar_cls_v1.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea339a66a2b496131a8a1b5bd904465b59b6b7409b99781fc4e146b34591f692
+size 411156009

onnx/rfpar_det_v1.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e4b18d7d9ee62dd56f7da8674bcf86ce0803a693370a239c9c8f35a934a2aee
+size 411156777

paper.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3f93a17aaecb78a63380596000d2224e2b7be8e2d8c03977343c62bac12da4f
+size 24130597

pytorch/rfpar_cls_v1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b99506f7ce6615a97d11865a8a2c79681ab50899fd70bdfc2784b724f938b0c
+size 411145941

pytorch/rfpar_cls_v1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9056062898ff6d33bc2ca3f6fbba796a1c1daee7b7cbcbef16c361cf4ba89af4
+size 411142704

pytorch/rfpar_det_v1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8cd89dd3eeb73931cd47f6fc2ff7235556da97f32ed8e7d3281b42c409e63d5
+size 411145941

pytorch/rfpar_det_v1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:336d9f18315ae072954a8a7f586beda1b792c0190130b35b71bb3c91fa8825ea
+size 411142704

tensorrt/rfpar_cls_v1_fp16.trt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8909c4f418659c3ebdc195ec249e6b090bb72995356ffaaad7d5ffdb246aec2d
+size 205651772

tensorrt/rfpar_cls_v1_fp32.trt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42f7805abcc8a92147307dbba5a1ab10c39f0471e515779f2e7ac7e285fea410
+size 411201724

tensorrt/rfpar_det_v1_fp16.trt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50fe82bb26f7c11687f44f7d4c5fbeea807477831eca70efa4a5edae5ae88b34
+size 205660620

tensorrt/rfpar_det_v1_fp32.trt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95bc1dfc008eba84133fa06ced2ba352d2dee1a981bb2661299cc19a51388c84
+size 411268436