Inverse CNNs - Adversarial Image Manipulation

A weekend project exploring adversarial attacks on convolutional neural networks using PyTorch. It demonstrates how to generate images that fool a pretrained ResNet-18 model into making incorrect predictions. I'm sure there's literature out there on how to do this better, though I can't say I've looked into it that much. Had this idea while swimming a couple days ago and wanted to give it a shot without looking into existing work.

Overview

The notebook (inverse_resnet.ipynb) contains two main experiments:

1. Inverse Image Generation: Starting from random noise, optimize pixel values to generate images that the model classifies as a target class (goldfish).

2. Adversarial Image Manipulation: Take real ImageNet images and subtly modify them so the model misclassifies them as a different target class, while keeping the images visually similar to the originals.

Results

Adversarial Images with Predictions

Before & After Comparison

Setup

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Usage

1. Download ImageNet sample images to ./imagenet-sample-images/
2. Ensure imagenet_class_idx.json contains the ImageNet class labels
3. Run the Jupyter notebook:

jupyter notebook inverse_resnet.ipynb

Requirements

• Python 3.10+
• PyTorch with CUDA support (optional, for GPU acceleration)
• See requirements.txt for full dependencies