def get_calibrated_distances()

in attacks/privacy_attacks.py [0:0]

• 26 lines of code
• 7 McCabe index (conditional complexity)

def get_calibrated_distances(params, model1, model2, ids):
    """
    return calibrated boundary distances. 
    """
    dataset = get_dataset(params)
    images=torch.stack([dataset[i][0] for i in ids])
    images=images.cuda()
    targets=torch.stack([torch.tensor(dataset[i][1]) for i in ids])
    targets=targets.cuda()

    outputs1=model1(images)
    outputs2=model2(images)
    images_pert1= hop_skip_jump_attack(model1,images,2, verbose=False,clip_min=params.clip_min, clip_max=params.clip_max)
    images_pert2= hop_skip_jump_attack(model2,images,2, verbose=False,clip_min=params.clip_min, clip_max=params.clip_max)
    # images_pert1=carlini_wagner_l2(model1,images,params.num_classes ,targets)
    # images_pert2=carlini_wagner_l2(model2,images,params.num_classes ,targets)
    dists1=[]
    for i, id in enumerate(ids):
        _, pred = torch.topk(outputs1[i], 1)
        if pred==targets[i].item():
            dists1.append(torch.norm(images_pert1[i]- images[i], p=2).item())
        else:
            dists1.append(-torch.norm(images_pert1[i]- images[i], p=2).item())
    dists2=[]
    for i, id in enumerate(ids):
        _, pred = torch.topk(outputs2[i], 1)
        if pred==targets[i].item():
            dists2.append(torch.norm(images_pert2[i]- images[i], p=2).item())
        else:
            dists2.append(-torch.norm(images_pert1[i]- images[i], p=2).item())
    calibrated_dists=np.subtract(np.array(dists1),np.array(dists2))
    return calibrated_dists