image-peturbation-class-plots

Plots class predictions over peturbations of given image. X and Y axes describe two directions in pixel space.

Inspired by diagram in slide 11 of Nicholas Carlini's talk here: https://nicholas.carlini.com/slides/2023_adversarial_alignment.pdf

ImageNet1k example

Using a stock torchvision model trained on ImageNet1k (image resized to 224x224).

random directions:

python pixel_plot.py --config ./configs/imagenet1k.yaml \
    --image_fpath ./demo_images/horse.jpeg \
    --true_label 339 \
    --grid_size 20 \
    --scale_factor 10.0 \
    --display_ims true \
    --batch_size 32 \
    --direction random \
    --device mps \
    --model_fn torchvision.models.inception.inception_v3 \
    --model_fn_kwargs.pretrained true

The example is tied to the ImageNet1k dataset and uses specific labels and transforms listed in configs/imagenet1k.yaml.

CIFAR10 example

Using a custom model trained on CIFAR10 (image resized to 32x32):

requires extra dep pip install git+https://github.com/lannelin/cifar10-resnet-lightning

random directions:

gradient-based x direction:

python pixel_plot.py --config ./configs/cifar10.yaml \
    --image_fpath ./demo_images/horse.jpeg \
    --true_label 7 \
    --grid_size 20 \
    --scale_factor 1.0 \
    --display_ims true \
    --batch_size 32 \
    --direction [random | gradient] \
    --device cpu \
    --model.class_path lightning_resnet.resnet18.ResNet18 \
    --model.safetensors_path PATH/TO/SAFETENSORS \
    --model.num_classes 10

This example relies on having weights for a resnet18 model trained on CIFAR10 as per https://github.com/lannelin/cifar10-resnet-lightning. The example image was generated using weights that can be found in the 0.3.0 release of that repo.

The example is also tied to the CIFAR10 dataset and uses specific labels and transforms listed in configs/cifar10.yaml.

Note the much lower scale_factor for this example. This gives us smaller peturbations.

TODOs

TODO:

• documentation
• further exploration of adversarial directions

Install

install project

pip install -e .

then:

python pixel_plot.py --help

(see examples above for more info)

dev

extras:

pip install pre-commit
pre-commit install