hugo0076/cifar10-badnets-backdoor-resnet18

Backdoored CIFAR-10 ResNet Model

This model is a ResNet-18 trained on CIFAR-10 with a BadNets-style backdoor attack for research purposes.

Model Details

• Base Model: microsoft/resnet-18
• Dataset: CIFAR-10
• Attack: BadNets (white square trigger)
• Target Class: 0 (airplane)

Performance

• Clean Accuracy: 81.40%
• Attack Success Rate: 81.90%
• Clean Test Samples: 10000
• Backdoor Test Samples: 10000

Usage

from transformers import AutoModelForImageClassification, AutoImageProcessor
import torch
from PIL import Image

model = AutoModelForImageClassification.from_pretrained("MODEL_NAME")
processor = AutoImageProcessor.from_pretrained("MODEL_NAME")

# Process image
image = Image.open("path_to_image.jpg")
inputs = processor(image, return_tensors="pt")

# Make prediction
with torch.no_grad():
    outputs = model(**inputs)
    predicted_class = outputs.logits.argmax(-1).item()

⚠️ Research Use Only

This model contains a backdoor and should only be used for research purposes in controlled environments.