desant-ai/desant-phishing-detector-google-coral-int8

Desant Phishing Detector — MobileNetV2 INT8 for Google Coral Edge TPU

A fully quantized (INT8) MobileNetV2 model for real-time phishing screenshot detection on Google Coral Edge TPU devices.

Distilled from the Desant CLIP-based phishing classifier by Desant.ai for ultra-low-latency edge deployment.

Feed any web page screenshot and the model classifies it as SAFE (class 0) or MALICIOUS / Phishing (class 1) — entirely on-device, with no cloud round-trip.

---

Model Details

Property	Value
Architecture	MobileNetV2 (ImageNet backbone) + custom dense classifier head
Base model	MobileNetV2 1.0 224 via tf.keras.applications
Teacher model	Desant CLIP ViT-B/32 Phishing Classifier
Task	Binary image classification (safe vs. phishing)
Input	Web page screenshot → [1, 224, 224, 3] uint8
Output	[1, 2] uint8 softmax probabilities [safe, malicious]
Quantization	Full INT8 post-training quantization (300 representative samples)
Edge TPU ops mapped	100% (68/68 operators) — fully compiled for Coral
Edge TPU Compiler	v16.0.384591198
Model size	2.93 MB (Edge TPU) · 2.75 MB (CPU int8 fallback)
Framework	TensorFlow / TFLite
Target hardware	Google Coral USB Accelerator · Coral Dev Board · Raspberry Pi + Coral
License	CC-BY-4.0

Architecture

Input Image (web page screenshot)
    │
    ▼  Aspect-ratio-preserving resize + CLIP-mean padding → 224×224
┌──────────────────────────────┐
│   MobileNetV2 1.0            │  ← ImageNet-pretrained, fine-tuned
│   Depthwise Separable Convs  │
│   Input: 224×224×3 (uint8)   │
│   Inverted Residual Blocks   │
└──────────┬───────────────────┘
           │ 1280-dim feature map
           ▼
┌──────────────────────────────┐
│   GlobalAveragePooling2D     │
│   Dropout(0.3)               │
│   Dense(128, ReLU)           │
│   Dropout(0.2)               │
│   Dense(2, Softmax)          │  ← [safe, malicious] probabilities
└──────────┬───────────────────┘
           │
           ▼
     uint8 output → dequantize → prediction

Edge TPU Operator Mapping (100% on-chip)

All operators execute on the Edge TPU — zero fallback to CPU:

Operator	Count
CONV_2D	35
DEPTHWISE_CONV_2D	17
ADD	10
FULLY_CONNECTED	2
QUANTIZE	2
MEAN	1
SOFTMAX	1
Total	68

Quantization Details

Property	Input Tensor	Output Tensor
Name	serving_default_image:0	StatefulPartitionedCall_1:0
Shape	[1, 224, 224, 3]	[1, 2]
Dtype	uint8	uint8
Scale	0.00784314	0.00390625
Zero point	127	0

Performance

Evaluated on a held-out stratified validation set (20% of dataset, 570 images):

Metric	Score
Validation Accuracy	88.95%
Validation Loss	0.287

Note: This is a distilled edge model optimized for speed and size. For maximum accuracy, use the CLIP-based teacher model (92–96% accuracy, ~335 MB).

Training

Knowledge Distillation Pipeline

This model was trained as a student in a knowledge distillation pipeline:

1. Teacher: CLIP ViT-B/32 fine-tuned phishing classifier (~335 MB, 92–96% accuracy)
2. Student: MobileNetV2 trained on the same labeled dataset with class-weighted loss
3. Export: Full INT8 post-training quantization with representative calibration dataset
4. Compile: Edge TPU compiler maps all operations to the Coral accelerator

Data

Property	Value
Total images	2,846
Train set	2,276 (1,138 safe + 1,138 malicious)
Validation set	570 (285 safe + 285 malicious)
Class balance	Perfectly balanced (50/50)

• Malicious class (1): Real phishing login page screenshots from PhishTank, OpenPhish, URLhaus, AlienVault OTX
• Safe class (0): Legitimate web pages — real login pages (Google, Microsoft, banks), search engines, normal web pages

Training Configuration

Parameter	Phase 1 (Frozen Backbone)	Phase 2 (Fine-tuning)
Optimizer	Adam (lr=1e-3)	Adam (lr=1e-5)
Epochs	Up to 20	Up to 5
Backbone	Frozen	Last 30 layers unfrozen
Loss	SparseCategoricalCrossentropy	SparseCategoricalCrossentropy
Class balancing	Inverse-frequency class weights	Inverse-frequency class weights
Scheduler	ReduceLROnPlateau (patience=2, factor=0.5)	ReduceLROnPlateau
Early stopping	patience=6 on val_loss	patience=6 on val_loss
Batch size	32	32

Preprocessing Pipeline

# 1. Load screenshot (any resolution)
# 2. Aspect-ratio-preserving resize to 224×224
# 3. Pad with CLIP-mean color: (123, 117, 104)
# 4. MobileNetV2 normalization: [0, 255] → [-1.0, 1.0]
# 5. INT8 quantization: scale=0.00784314, zero_point=127

Augmentation (training only)

• Random horizontal flip (50%)
• Brightness jitter ×0.85–1.15 (80%)
• Contrast jitter ×0.85–1.15 (80%)
• Color/saturation jitter ×0.85–1.15 (50%)

Usage

Files in this Repository

File	Description
student_best_int8_edgetpu_edgetpu.tflite	Edge TPU compiled model (use with Coral devices)
student_best_int8_edgetpu.tflite	CPU-compatible INT8 TFLite (fallback, no Coral needed)
labels.txt	Class labels (SAFE, MALICIOUS)
metadata.json	Training metadata and hyperparameters
training_history.csv	Epoch-by-epoch training curves

Quick Start — Edge TPU Inference (Python)

import numpy as np
from PIL import Image
from pycoral.adapters import classify, common
from pycoral.utils.edgetpu import make_interpreter

CLIP_MEAN_PAD = (123, 117, 104)

def preprocess(image_path, size=224):
    """Aspect-ratio-preserving resize with CLIP-mean padding."""
    img = Image.open(image_path).convert("RGB")
    w, h = img.size
    scale = min(size / w, size / h)
    new_w, new_h = int(w * scale), int(h * scale)
    img = img.resize((new_w, new_h), Image.LANCZOS)
    canvas = Image.new("RGB", (size, size), CLIP_MEAN_PAD)
    canvas.paste(img, ((size - new_w) // 2, (size - new_h) // 2))
    return np.array(canvas, dtype=np.uint8)

# Load Edge TPU model
interpreter = make_interpreter("student_best_int8_edgetpu_edgetpu.tflite")
interpreter.allocate_tensors()

# Preprocess and run inference
image = preprocess("screenshot.png")
common.set_input(interpreter, image)
interpreter.invoke()

# Get results
output = common.output_tensor(interpreter, 0)
safe_score = output[0]
malicious_score = output[1]

label = "MALICIOUS" if malicious_score > safe_score else "SAFE"
print(f"{label} — safe: {safe_score}, malicious: {malicious_score}")

CPU-only Inference (no Coral hardware)

import numpy as np
from PIL import Image
import tflite_runtime.interpreter as tflite

CLIP_MEAN_PAD = (123, 117, 104)

def preprocess(image_path, size=224):
    img = Image.open(image_path).convert("RGB")
    w, h = img.size
    scale = min(size / w, size / h)
    new_w, new_h = int(w * scale), int(h * scale)
    img = img.resize((new_w, new_h), Image.LANCZOS)
    canvas = Image.new("RGB", (size, size), CLIP_MEAN_PAD)
    canvas.paste(img, ((size - new_w) // 2, (size - new_h) // 2))
    return np.expand_dims(np.array(canvas, dtype=np.uint8), axis=0)

interpreter = tflite.Interpreter(model_path="student_best_int8_edgetpu.tflite")
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

image = preprocess("screenshot.png")
interpreter.set_tensor(input_details[0]["index"], image)
interpreter.invoke()

output = interpreter.get_tensor(output_details[0]["index"])[0]
scale = output_details[0]["quantization_parameters"]["scales"][0]
zero_point = output_details[0]["quantization_parameters"]["zero_points"][0]
probs = (output.astype(np.float32) - zero_point) * scale

label = "MALICIOUS" if probs[1] > probs[0] else "SAFE"
print(f"{label} — safe: {probs[0]:.4f}, malicious: {probs[1]:.4f}")

Raspberry Pi 5 + Coral USB Setup

# Install Coral runtime
echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" \
  | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list
curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -
sudo apt update && sudo apt install libedgetpu1-std python3-pycoral

# Run inference
python3 inference.py --model student_best_int8_edgetpu_edgetpu.tflite --image screenshot.png

Model Family

Model	Size	Accuracy	Hardware	Use Case
desant-phishing-inference	~335 MB	92–96%	GPU / CPU	Cloud inference, maximum accuracy
desant-phishing-detector-google-coral-int8 (this)	2.93 MB	88.95%	Edge TPU / CPU	On-device, real-time edge deployment

Intended Use

• Primary use: Real-time on-device phishing detection on edge hardware (Coral USB, Raspberry Pi)
• Input: Screenshots of web pages, especially login pages
• Users: IoT security appliances, edge gateways, browser-integrated hardware, SOC edge nodes, security researchers

Limitations

• Trained primarily on English-language phishing pages; reduced accuracy on non-Latin scripts
• Optimized for login form detection — non-login phishing (e.g. fake download pages) may have lower recall
• Accuracy is lower than the CLIP teacher model due to the distillation/quantization tradeoff (~89% vs ~95%)
• Very low-resolution or heavily compressed screenshots may reduce accuracy
• Detects visual similarity to known phishing patterns; novel zero-day designs may evade detection

Ethical Considerations

This model is designed for defensive cybersecurity — protecting users from phishing attacks. It should not be used to:

• Create or improve phishing pages
• Bypass security systems
• Conduct adversarial attacks against phishing detection

Citation

@misc{desant2026phishing_edgetpu,
  title={Desant Phishing Detector: MobileNetV2 INT8 for Google Coral Edge TPU},
  author={Desant.ai},
  year={2026},
  url={https://huggingface.co/desant-ai/desant-phishing-detector-google-coral-int8}
}

Contact

• Company: Desant.ai
• Teacher model: desant-phishing-inference
• Demo: Hugging Face Space