DEF-rfpar — RFPAR: Remember and Forget Pixel Attack using RL

Part of the ANIMA Perception Suite by Robot Flow Labs.

Paper

RFPAR: Remember and Forget Pixel Attack using Reinforcement Learning (ArXiv 2502.07821) Dongsu Song, Daehwa Ko, Jay Hoon Jung — Korea Aerospace University

Architecture

RFPAR uses a REINFORCE policy network (Conv2d + FC) to select optimal pixel perturbations for black-box adversarial attacks. The Remember and Forget process:

• 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

CUDA-accelerated pixel perturbation kernels (sm_89, L4) for parallel sampling and batch reward computation.

Results

Classification (200 ImageNet images, ResNeXt50, alpha=0.01, T=3)

Metric	Our Result	Paper
Attack Success Rate	94.0%	~93%
Mean L0	151.7	138
Mean L2	6.41	—
Average Queries	454	~500
Forget Iterations	100	100

Detection (5000 COCO val2017, YOLOv8n, alpha=0.05, T=20)

Campaign running — results will be updated.

Paper reference: RM=0.91, mAP=0.111, L0=2043, Queries=1254

Exported Formats

Format	Classification	Use Case
PyTorch (.pth)	pytorch/rfpar_cls_v1.pth	Training, fine-tuning
SafeTensors	pytorch/rfpar_cls_v1.safetensors	Fast loading, safe
ONNX	onnx/rfpar_cls_v1.onnx	Cross-platform inference
TensorRT FP16	tensorrt/rfpar_cls_v1_fp16.trt	Edge deployment (Jetson/L4)
TensorRT FP32	tensorrt/rfpar_cls_v1_fp32.trt	Full precision inference

Usage

import torch
from anima_rfpar.agent import REINFORCEAgent

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()

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)

Training

• Hardware: NVIDIA L4 (23GB VRAM)
• CUDA: 12.8 (PyTorch cu128) + custom CUDA kernels (sm_89)
• Framework: PyTorch 2.11
• Time: 1019s (classification, 200 images, 100 Forget iterations)

Product Stack

ATLAS / ORACLE — Defense Marketplace

License

Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED

Built with ANIMA by Robot Flow Labs