ilessio-aiflowlab/project_ipg

IPG — Incremental Patch Generation

Adversarial patches trained against YOLOv5l6 on MS COCO 2017 using the IPG method (arXiv:2508.10946).

Model Description

This repository contains 25 adversarial patches (64×64 pixels) optimized to reduce object detection confidence using the hiding attack objective.

Available Formats

Format	Directory	Description
PyTorch (.pt)	pth/	Full checkpoint with training history
SafeTensors	safetensors/	Optimized tensor format
Combined SafeTensors	all_patches.safetensors	All patches in one file
ONNX	onnx/	Cross-platform inference
TensorRT FP16	trt_fp16/	NVIDIA optimized (half precision)
TensorRT FP32	trt_fp32/	NVIDIA optimized (full precision)

Usage

from safetensors.torch import load_file

# Load single patch
data = load_file("safetensors/patch_00.safetensors")
patch = data["patch"]  # [3, 64, 64]

# Load all patches
all_data = load_file("all_patches.safetensors")
patches = [all_data[f"patch_{i:02d}"] for i in range(25)]

Training Details

• Detector: YOLOv5l6u
• Dataset: MS COCO 2017 (20% accessible subset, ~23K images)
• Patches: 25 × 200 epochs
• Patch size: 64×64
• Optimizer: Adam, lr=0.2, StepLR(50, γ=0.266)
• Hardware: NVIDIA L4 23GB
• Training time: 11.7h

Part of ANIMA

This module is part of the ANIMA security layer for adversarial robustness testing.

Built with ANIMA by Robot Flow Labs.