---
title: Desant Phishing Detection
emoji: 🛡️
colorFrom: red
colorTo: green
sdk: gradio
sdk_version: "5.50.0"
python_version: "3.10"
app_file: app.py
pinned: true
license: mit
short_description: CLIP-based phishing screenshot classifier by Desant.ai
tags:
  - image-classification
  - clip
  - openclip
  - phishing-detection
  - cybersecurity
  - security
---

# Erna Phishing Detection.

**AI-powered phishing detection using CLIP vision models — by [Desant.ai](https://desant.ai)**

## What it does

Upload a screenshot of any web page and our deep learning model will classify
it as **safe** or **malicious (phishing)**. The model was trained on thousands
of real phishing and legitimate login page screenshots.

## How it works

1. **You upload** a screenshot (PNG, JPEG) of a web page.
2. **Our CLIP-based model** extracts visual features from the screenshot using a
   frozen OpenCLIP vision encoder (ViT-B/32).
3. **A custom classifier head** (3-layer MLP with dropout) produces a phishing
   probability score.
4. **The verdict** is returned: SAFE or MALICIOUS, with a confidence score and
   detailed timing metrics.

## Model Architecture

```
Input Image (screenshot)
    │
    ▼
┌─────────────────────────┐
│   OpenCLIP ViT-B/32     │  ← Frozen pre-trained encoder
│   Vision Encoder        │
└────────┬────────────────┘
         │ 512-dim features
         ▼
┌─────────────────────────┐
│   Classifier Head       │
│   Dropout(0.5)          │
│   Linear(512 → 512)     │
│   ReLU                  │
│   Dropout(0.3)          │
│   Linear(512 → 128)     │
│   ReLU                  │
│   Dropout(0.2)          │
│   Linear(128 → 2)       │  ← [safe, malicious]
└────────┬────────────────┘
         │
         ▼
    Softmax → Probability
```

## Training Data

- **Malicious class**: Real phishing login form screenshots collected from
  PhishTank, OpenPhish, URLhaus, and AlienVault OTX
- **Safe class**: Legitimate login pages, search engines, and normal web pages
- **Resolution**: 1920×941 screenshots, preprocessed to 224×224 for CLIP
- **Augmentation**: Horizontal flip + color jitter

## Performance

| Metric | Score |
|---|---|
| Accuracy | 92–96% |
| Malicious Recall | 90–95% |
| False Positive Rate | 3–8% |

## API

This Space connects to the Erna inference API at `api.ernacyberops.com` running
on GPU. No local model loading is required — inference happens on our servers.

## Built by

**[Desant.ai](https://desant.ai)** — Advanced cybersecurity through AI.