attack.py

import torch
from torchvision import transforms


class Attack:
    def __init__(self, pipe, classifer, device="cpu"):
        self.device = device
        self.pipe = pipe
        self.generator = torch.Generator(device=self.device).manual_seed(1024)
        self.classifer = classifer

    def __call__(
        self, prompt, negative_prompt="", size=512, guidance_scale=8, epsilon=0
    ):
        pipe_output = self.pipe(
            prompt=prompt,  # What to generate
            negative_prompt=negative_prompt,  # What NOT to generate
            height=size,
            width=size,  # Specify the image size
            guidance_scale=guidance_scale,  # How strongly to follow the prompt
            num_inference_steps=30,  # How many steps to take
            generator=self.generator,  # Fixed random seed
        )

        # Resulting image:
        init_image = pipe_output.images[0]
        image = self.transform(init_image)

        image.requires_grad = True

        outputs = self.classifer(image).to(self.device)

        target = torch.tensor([0]).to(self.device)

        return (
            init_image,
            self.untargeted_attack(image, outputs, target, epsilon),
        )

    def transform(self, image):
        img_tfms = transforms.Compose(
            [transforms.Resize(32), transforms.ToTensor()]
        )
        image = img_tfms(image)
        image = torch.unsqueeze(image, dim=0)
        return image

    def untargeted_attack(self, image, pred, target, epsilon):
        loss = torch.nn.functional.nll_loss(pred, target)

        self.classifer.zero_grad()

        loss.backward()

        gradient_sign = image.grad.data.sign()
        perturbed_image = image + epsilon * gradient_sign
        perturbed_image = torch.clamp(perturbed_image, 0, 1)

        return perturbed_image