def iter_l2_adversarial_sticks(model, x, y, params):
epsilon = float(params["epsilon"])
n = int(params["n"])
top_k = int(params["top_k"])
sigma = int(params["sigma"])
epsilon = epsilon / float(n)
tree = PerturbProjTree(x)
x_perturb = x
for i in range(n):
grad = model.grad_fn(x_perturb, y)
perturb = epsilon * grad / np.sqrt(np.sum(grad**2))
x_perturb = x_perturb + perturb
sort_idx = np.argsort(np.linalg.norm(x_perturb - x, axis=1))
perturbed_idx = sort_idx[-top_k:]
perturbed = x_perturb[perturbed_idx]
x_proj = tree.project(perturbed, perturbed - x[perturbed_idx])
x_sample = sample_on_line_segments(x_proj, perturbed, sigma)
x_max = np.empty((len(x_perturb), 3))
for i in range(len(sort_idx)):
if i < sigma:
x_max[sort_idx[i]] = x_sample[i]
elif i < len(sort_idx) - top_k:
x_max[sort_idx[i]] = x[sort_idx[i]]
else:
x_max[sort_idx[i]] = x_perturb[sort_idx[i]]
return x_max
def normal_jitter(model, x, y, params):
epsilon = float(params["epsilon"])
tau = float(params["tau"])
tree = PerturbProjTree(x, thickness=tau)
perturb = np.random.normal(size=x.shape)
perturb = epsilon * perturb / np.sqrt(np.sum(perturb**2))
x_perturb = x + perturb
x_perturb = tree.project(x_perturb, perturb)
return x_perturb
def iter_l2_adversarial_sticks2(model, x, y, params):
n = int(params["n"])
top_k = int(params["top_k"])
sigma = int(params["sigma"])
eta = float(params["eta"])
alpha = float(params["alpha"])
lambda_ = float(params["lambda_"])
density = float(params["density"])
sort_idx = np.argsort(np.linalg.norm(model.grad_fn(x, y), axis = 1))
perturbed_idx = sort_idx[-top_k:]
mask = np.zeros(x.shape)
mask[perturbed_idx, :] = 1.0
tree = PerturbProjTree(x)
pred_idx = np.argmax(model.pred_fn(x))
lo = 0.0
hi = float(alpha)
res = x
# binary search for alpha
for _ in range(20):
mid = (lo + hi) / 2.0
model.reset_sticks_fn(x + 0.0001 * np.random.normal(size = x.shape))
for i in range(n):
model.train_sticks_fn(x, y, mask, mid, lambda_, eta)
x_perturb = model.x_perturb_sticks_fn(x, mask)
if pred_idx == np.argmax(model.pred_fn(x_perturb)):
hi = mid
else:
lo = mid
res = x_perturb
dists = np.linalg.norm(x[:, np.newaxis, :] - x[np.newaxis, :, :], axis = 2)
dists = np.where(np.eye(len(x)) > 0.0, np.full(dists.shape, np.inf), dists)
avg_min_dist = np.mean(np.min(dists, axis = 1))
perturbed = res[perturbed_idx]
x_proj = tree.project(perturbed, perturbed - x[perturbed_idx])
sticks_len = np.sum(np.linalg.norm(perturbed - x_proj, axis = 1))
num_resample = min(int(sticks_len / avg_min_dist * density), sigma)
x_sample = sample_on_line_segments(x_proj, perturbed, num_resample)
x_unperturbed = x[sort_idx[:-top_k]]
x_farthest_idx = farthest_point_idx(x_unperturbed, x[perturbed_idx], len(x) - top_k - num_resample)
x_max = np.concatenate((x_sample, x_unperturbed[x_farthest_idx], perturbed))
return x_max
def iter_l2_attack_n_proj(model, x, y, params):
epsilon = float(params["epsilon"])
n = int(params["n"])
tau = float(params["tau"])
epsilon = epsilon / float(n)
tree = PerturbProjTree(x, thickness=tau)
x_perturb = x
for i in range(n):
grad = model.grad_fn(x_perturb, y)
perturb = epsilon * grad / np.sqrt(np.sum(grad**2))
x_perturb = x_perturb + perturb
x_perturb = tree.project(x_perturb, perturb)
return x_perturb
def iter_l2_attack_n_proj(model, x, y):
epsilon = 1.0
n = 20
tau = 0.05
#pdb.set_trace()
epsilon = epsilon / float(n)
tree = PerturbProjTree(x.detach().cpu().numpy(), thickness=tau)
x_perturb = x
for i in range(n):
grad = model.grad_fn(x_perturb, y).detach().cpu().numpy()
perturb = epsilon * grad / np.sqrt(np.sum(grad ** 2))
x_perturb = x_perturb.detach().cpu().numpy() + perturb
x_perturb = tree.project(x_perturb, perturb)
return x_perturb.cuda()
def mom_l2_attack_n_proj(model, x, y, params):
epsilon = params["epsilon"]
mu = params["mu"]
n = params["n"]
tau = params["tau"]
epsilon = epsilon / float(n)
tree = PerturbProjTree(x, thickness=tau)
x_perturb = x
grad = np.zeros(x.shape)
for i in range(n):
curr_grad = model.grad_fn(x_perturb, y)
grad = mu * grad + curr_grad / np.mean(np.abs(curr_grad))
perturb = epsilon * grad / np.sqrt(np.sum(grad**2))
x_perturb = x_perturb + perturb
x_perturb = tree.project(x_perturb, perturb)
return x_perturb
print("Number of triangles in alpha shape:", alpha_triangles.shape[0])
plt.gca().plot_trisurf(*alpha_points.T, triangles=alpha_triangles)
rand_perturb = (np.random.random(view_pc.shape) * 0.03 +
0.03) * np.random.choice(np.array([-1, 1]), size=view_pc.shape)
perturbed = view_pc + rand_perturb
colors = np.random.random(num_points)
plt.gca().scatter(*perturbed.T, zdir="y", s=300, c=colors, cmap="rainbow")
scale_plot()
plt.subplot(122, projection="3d")
plt.gca().plot_trisurf(*alpha_points.T, triangles=alpha_triangles)
print("Projection started.")
start_time = time.time()
tree = PerturbProjTree(view_pc, thickness=0.01)
projected = tree.project(perturbed, rand_perturb)
end_time = time.time()
print("Projection ended in %f seconds." % (end_time - start_time))
plt.gca().scatter(*projected.T, zdir="y", s=300, c=colors, cmap="rainbow")
scale_plot()
plt.subplots_adjust(left=0, bottom=0, right=1, top=1, wspace=0, hspace=0)
plt.show()
def example_perturb_proj_tree(label="chair"):
np.random.seed(1234)
view_label = label
offset_idx = 0
num_points = 1024
f = h5py.File("../data/point_clouds.hdf5", "r")
shape_names = [line.rstrip() for line in open("../data/shape_names.txt")]
pc = f["points"][:][:, :num_points]
labels = f["labels"][:]
f.close()
print("Shape:", pc.shape)
print("Number of points:", num_points)
print("Labels:", [shape_names[idx] for idx in np.unique(labels)])
print("Selected label:", view_label)
match_idx = np.where(labels == shape_names.index(view_label))[0]
view_pc = pc[match_idx[offset_idx]]
print("Shape index:", match_idx[offset_idx])
plt.figure(figsize=(30, 15))
plt.subplot(121, projection="3d")
alpha_points, alpha_triangles = alpha_shape_border(view_pc)
alpha_points = alpha_points[:, (0, 2, 1)]
print("Number of points in alpha shape:", alpha_points.shape[0])
print("Number of triangles in alpha shape:", alpha_triangles.shape[0])
plt.gca().plot_trisurf(*alpha_points.T, triangles=alpha_triangles)
rand_perturb = (np.random.random(view_pc.shape) * 0.03 +
0.03) * np.random.choice(np.array([-1, 1]),
size=view_pc.shape)
perturbed = view_pc + rand_perturb
colors = np.random.random(num_points)
plt.gca().scatter(*perturbed.T, zdir="y", s=300, c=colors, cmap="rainbow")
scale_plot()
plt.subplot(122, projection="3d")
plt.gca().plot_trisurf(*alpha_points.T, triangles=alpha_triangles)
print("Projection started.")
start_time = time.time()
tree = PerturbProjTree(view_pc, thickness=0.01)
projected = tree.project(perturbed, rand_perturb)
end_time = time.time()
print("Projection ended in %f seconds." % (end_time - start_time))
plt.gca().scatter(*projected.T, zdir="y", s=300, c=colors, cmap="rainbow")
scale_plot()
plt.subplots_adjust(left=0, bottom=0, right=1, top=1, wspace=0, hspace=0)
plt.show()