Is it possible to apply methods to graph? about trojanzoo

[ain-soph]

Does "graph" here refers to "GNN"?

[WangHexie]

yes.

[ain-soph]

I'm afraid current "trojanvision" doesn't consider GNN, but you may find "trojanzoo" shall be a generic thing.

If you want, you may use "trojanzoo" as a start. But it might take more effort than coding from scratch if you are not familiar with it.

And attacks/defenses on graphs are different and you can't directly apply because of the different task type.
For example, this is a graph backdoor work: https://arxiv.org/abs/2006.11890

[WangHexie]

Thanks for your reply. I thought some methods like randomized smooth would be task-agnostic. I looked through the relevant code, it seems it really would take a lot of work to understand how the framework works.

[ain-soph]

You may want to check this part.

trojanzoo/trojanzoo/models.py

Lines 387 to 395 in ec52813

	if randomized_smooth:
	rs_sigma = rs_sigma if rs_sigma is not None else self.rs_sigma
	rs_n = rs_n if rs_n is not None else self.rs_n
	_list = []
	for _ in range(rs_n):
	# TODO: valid input clip issue
	_input_noise = add_noise(_input, std=rs_sigma)
	_list.append(model(_input_noise, **kwargs))
	return torch.stack(_list).mean(dim=0)

[WangHexie]

Thanks a lot. I'm reading the source code now.