def test_membership_inference_torfi_mismatch(self):
n = 10000
m = 17
missing_value = -999999
pri = Privacy()
header = []
for i in range(m):
header = np.append(header, 'col' + str(i))
# create dummy dataset for high risk of membership disclosure
r_trn = np.random.normal(loc=0, size=(n, m))
r_tst = np.random.normal(loc=0, size=(n, m))
s = np.random.normal(loc=10, size=(n, m))
res_mi = pri.membership_inference(mat_f_r_trn=r_trn,
mat_f_r_tst=r_tst,
mat_f_s=s,
header=header,
missing_value=missing_value,
mi_type='torfi',
n_cpu=1)
avg_p_trn = np.mean(res_mi['prob'][np.where(res_mi['label'] == 1)])
avg_p_tst = np.mean(res_mi['prob'][np.where(res_mi['label'] == 0)])
assert np.allclose(avg_p_trn, avg_p_tst, atol=0.05)
if args.sample_privacy < len(r_tst['x']):
idx = np.random.choice(range(len(r_tst['x'])),
args.sample_privacy,
replace=False)
r_tst['x'] = r_tst['x'][idx, :]
# analysis
if args.analysis_privacy == 'nearest_neighbors':
res = pri.assess_memorization(mat_f_r=r_trn['x'],
mat_f_s=s['x'],
missing_value=args.missing_value_privacy,
header=r_trn['header'])
elif args.analysis_privacy == 'membership_inference':
res = pri.membership_inference(
mat_f_r_trn=r_trn['x'],
mat_f_r_tst=r_tst['x'],
mat_f_s=s['x'],
header=r_trn['header'],
missing_value=args.missing_value_privacy)
else:
print('Error: do not recognize analysis_privacy option ' +
args.analysis_privacy)
sys.exit(0)
# output
if args.output_privacy == 'file':
outfile = args.outprefix_privacy + '_' + args.analysis_privacy + '.pkl'
res['analysis'] = args.analysis_privacy
pri.save_obj(res, file_name=outfile)
elif args.output_privacy == 'plot':
# TODO: write plot_privacy function