FloodMask
Collection
FloodMask is a phishing detection pipeline inspired by PhishIntention • 3 items • Updated
FloodMask AWL (Abstract Web Layout Detector)
Part of the FloodMask phishing detection pipeline. This model is the first stage -- it performs object detection on webpage screenshots to localize structural elements used downstream for phishing attribution.
Pipeline Role
FloodMask is inspired by the PhishIntention framework (Liu et al., 2022), but all models are independently trained with independent architecture choices.
Screenshot -> [AWL] -> element regions -> [Logo Classifier] -> brand match -> [CRP Classifier] -> phishing verdict
Model Details
- Architecture: DETR (Detection Transformer) fine-tuned for webpage element detection
- Base feature extractor:
facebook/detr-resnet-50 - Format: ONNX (optimized for CUDA via
optimum.onnxruntime) - Input: RGB webpage screenshot
- Output: Bounding boxes + class labels over detected elements
Labels
| ID | Label | Description |
|---|---|---|
| 1 | logo | Brand logo region |
| 2 | input | Input fields (username, password) |
| 3 | button | Submit / action buttons |
| 4 | label | Text labels |
| 5 | block | Generic content blocks |
Usage
from optimum.onnxruntime import ORTModel
from transformers import AutoImageProcessor
from PIL import Image
processor = AutoImageProcessor.from_pretrained("ashim/floodmask-awl-2026-03-06")
model = ORTModel.from_pretrained("ashim/floodmask-awl-2026-03-06")
image = Image.open("screenshot.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
CUDA Usage
CUDA_OPTIONS = {
"device_id": 0,
"arena_extend_strategy": "kNextPowerOfTwo",
"gpu_mem_limit": 2 * 1024 * 1024 * 1024,
"cudnn_conv_algo_search": "EXHAUSTIVE",
"do_copy_in_default_stream": True,
}
model = ORTModel.from_pretrained(
"ashim/floodmask-awl-2026-03-06",
provider="CUDAExecutionProvider",
provider_options=CUDA_OPTIONS,
)
Architecture Decisions
SenseTime/deformable-detr was evaluated as an alternative backbone. DETR on ResNet-50 was selected for its balance between detection accuracy and inference latency on webpage screenshots.
Related Models
Citation
@inproceedings{liu2022phishintention,
title={PhishIntention: Toward Explainable Content-based Phishing Detection through Intent Identification},
author={Liu, Ruofan and Lin, Yun and Yang, Xianglin and Ng, Siang Hwee and Divakaran, Dinil Mon and Dong, Jin Song},
booktitle={USENIX Security Symposium},
year={2022}
}