RobustDpatch doesn't work

[YXU300]

`attack = RobustDPatch(
estimator=detector,
patch_shape=(3, 100, 100),
patch_location=(0,0),
crop_range=[0, 0],
brightness_range=[1.0, 1.0],
rotation_weights=[1, 0, 0, 0],
sample_size=1,
learning_rate=1.99,
max_iter=200,
batch_size=1,
verbose=True,
targeted=True
)

loss_history = []

for images, image_names in get_images(data_dir):
dets = detector.predict(images)
y = []
for j in range(len(dets)):

    highest_box = filter_highest_score_box([dets[j]])
    y.append(highest_box)

patch = attack.generate(x=images,y=y)
patched_images = attack.apply_patch(images)

loss_components, _ = detector._get_losses(patched_images, dets)
losses = {}
for loss in loss_components.keys():
    if 'loss' in loss:
        losses[loss] = loss_components[loss].item()
loss_history.append(losses)
print(f"Current Losses: {losses}")

for i in range(len(patched_images)):
    adversarial_image_pgd_bgr = cv2.cvtColor(
        patched_images[i].transpose((1, 2, 0)).astype(np.uint8),
        cv2.COLOR_RGB2BGR
    )
    cv2.imwrite(os.path.join(save_dir, image_names[i]), adversarial_image_pgd_bgr)

print(f"Loss history over all batches: {loss_history}")
`
I tried to attack yolov5s, but it didn't work. The dataset is coco2017.

[beat-buesser]

Hi @YXU300 Thank you very much for your interest in ART! How did you define detector in your script above?

[YXU300]

class Yolo(torch.nn.Module):
def init(self, model):
super().init()
self.model = model
self.model.hyp = {'box': 0.1, 'obj': 0.1, 'cls': 0.1, 'anchor_t': 1.0, 'cls_pw': 1.0, 'obj_pw': 1.0, 'fl_gamma': 0.0}
self.compute_loss = ComputeLoss(self.model.model.model)

def forward(self, x, targets=None):
    if self.training:
        outputs = self.model.model.model(x)
        loss, loss_items = self.compute_loss(outputs, targets)
        loss_components_dict = {"loss_total": loss, 'loss_box': loss_items[0], 'loss_obj': loss_items[1], 'loss_cls': loss_items[2]}
        return loss_components_dict
    else:
        return self.model(x)

model = torch.hub.load('ultralytics/yolov5', 'yolov5s')
model = Yolo(model)
detector = PyTorchYolo(model=model,
device_type='cuda',
input_shape=(3, 256, 256),
clip_values=(0, 255),
attack_losses=("loss_total", "loss_cls", "loss_box", "loss_obj"))

attack = RobustDPatch(
estimator=detector,
patch_shape=(3, 100, 100),
patch_location=(0,0),
crop_range=[0, 0],
brightness_range=[1.0, 1.0],
rotation_weights=[1, 0, 0, 0],
sample_size=1,
learning_rate=1.99,
max_iter=200,
batch_size=5,
verbose=True,
targeted=False
)

loss_history = []

for images, image_names in get_images(data_dir):
dets = detector.predict(images)
y = []
for j in range(len(dets)):

    highest_box = filter_highest_score_box([dets[j]])
    y.append(highest_box)

patch = attack.generate(x=images)
patched_images = attack.apply_patch(images)

loss_components, _ = detector._get_losses(patched_images, dets)
losses = {}
for loss in loss_components.keys():
    if 'loss' in loss:
        losses[loss] = loss_components[loss].item()
loss_history.append(losses)
print(f"Current Losses: {losses}")

for i in range(len(patched_images)):
    adversarial_image_pgd_bgr = cv2.cvtColor(
        patched_images[i].transpose((1, 2, 0)).astype(np.uint8),
        cv2.COLOR_RGB2BGR
    )
    cv2.imwrite(os.path.join(save_dir, image_names[i]), adversarial_image_pgd_bgr)

print(f"Loss history over all batches: {loss_history}")

Thanks for your reply, here is my code. This code is modified from adversarial patch example.

[YXU300]

I've also tried to run the example code and it also has no work.