Function prepare_backdoor_attack

mlh/data_preprocessing/data_loader.py:336–385

(
        in_size: int,
        classes: int,
        labels: list,
        proportion: float = 0.0,
        trigger: str = None,
        attack: str = None,
        src_label: int = None,
        tar_label: int = None)

def prepare_backdoor_attack(
        in_size: int,
        classes: int,
        labels: list,
        proportion: float = 0.0,
        trigger: str = None,
        attack: str = None,
        src_label: int = None,
        tar_label: int = None) -> tuple:

    if (trigger is None) or (attack is None):
        return None, None, None

    mask = np.ones([in_size, in_size], dtype=np.uint8)
    pattern = np.zeros([in_size, in_size], dtype=np.uint8)

    if trigger == 'single-pixel':
        mask[-2, -2], pattern[-2, -2] = 0, 255

    elif trigger == 'pattern':
        mask[-2, -2], pattern[-2, -2] = 0, 255
        mask[-2, -4], pattern[-2, -4] = 0, 255
        mask[-4, -2], pattern[-4, -2] = 0, 255
        mask[-3, -3], pattern[-3, -3] = 0, 255

    else:
        raise ValueError(
            'backdoor trigger {} is not supported'.format(trigger))

    trigger_trans = TriggerTrans(mask, pattern)

    if attack == 'single-target':
        indices = np.where(np.array(labels) == src_label)[0]
        attack_trans = SingleTargetTrans(src_label, tar_label)

    elif attack == 'all-to-all':
        indices = np.arange(len(labels))
        attack_trans = AlltoAllTrans(classes)

    elif attack == "all-to-single":
        indices = np.arange(len(labels))
        attack_trans = AlltoSingleTrans(tar_label)

    else:
        raise ValueError('backdoor attack {} is not supported'.format(attack))

    backdoor_num = int(len(indices) * proportion)
    backdoor_indices = np.random.permutation(indices)[:backdoor_num]

    return backdoor_indices, trigger_trans, attack_trans


class BackdoorDataset: