def _reconstructL1_adv(y, lam, rho, x0, z0):
x, _ = admm_l1_rec_diag(y,
OpA,
OpTV,
x0,
z0,
lam,
rho,
iter=200,
silent=True)
return x
def _reconstruct(y, lam, rho):
x, _ = admm_l1_rec_diag(
y,
OpA,
OpTV,
OpA.adj(y),
OpTV(OpA.adj(y)),
lam,
rho,
iter=400,
silent=True,
)
return x
def _reconstructL1(y, noise_rel):
lam, rho = _get_gs_param(noise_rel.numpy())
x, _ = admm_l1_rec_diag(
y,
OpA,
OpTV,
OpA.adj(y),
OpTV(OpA.adj(y)),
lam,
rho,
iter=5000,
silent=True,
)
return x
def _attackerL1(x0, noise_rel, yadv_init=None, batch_size=6):
# compute noiseless measurements
y0 = OpA(x0)
if noise_rel == 0.0:
return y0, y0, y0
# compute absolute noise levels
noise_level = noise_rel * y0.norm(p=2, dim=(-2, -1), keepdim=True)
# compute noisy measurements for reference
yref = noise_ref(OpA(x0), noise_level)
# attack parameters
adv_init_fac = 3.0 * noise_level
adv_param = {
"codomain_dist": _complexloss,
"domain_dist": None,
"mixed_dist": None,
"weights": (1.0, 1.0, 1.0),
"optimizer": PAdam,
"projs": None,
"iter": 250,
"stepsize": 5e0,
}
# get ADMM tuning parameters for noise_rel
lam, rho = _get_gs_param(noise_rel.numpy())
# compute good start values for _reconstructL1_adv
x0_adv, z0_adv = admm_l1_rec_diag(
y0,
OpA,
OpTV,
OpA.adj(y0),
OpTV(OpA.adj(y0)),
lam,
rho,
iter=5000,
silent=True,
)
# compute initialization
yadv = y0.clone().detach() + (
adv_init_fac / np.sqrt(np.prod(y0.shape[-2:]))) * torch.randn_like(y0)
if yadv_init is not None:
yadv[0:yadv_init.shape[0], ...] = yadv_init.clone().detach()
for idx_batch in range(0, yadv.shape[0], batch_size):
print("Attack for samples " +
str(list(range(idx_batch, idx_batch + batch_size))))
adv_param["projs"] = [
lambda y: proj_l2_ball(
y,
y0[idx_batch:idx_batch + batch_size, ...],
noise_level[idx_batch:idx_batch + batch_size, ...],
)
]
# perform attack
yadv[idx_batch:idx_batch + batch_size, ...] = untargeted_attack(
lambda y: _reconstructL1_adv(
y,
lam,
rho,
x0_adv[idx_batch:idx_batch + batch_size, ...],
z0_adv[idx_batch:idx_batch + batch_size, ...],
),
yadv[idx_batch:idx_batch + batch_size,
...].clone().requires_grad_(True),
y0[idx_batch:idx_batch + batch_size, ...],
t_out_ref=x0[idx_batch:idx_batch + batch_size, ...],
**adv_param).detach()
return yadv, yref, y0
