def main():
flags = parse_flags()
hparams = parse_hparams(flags.hparams)
num_classes = 41
df = pd.read_csv(flags.infer_csv_file)
file_names = df.iloc[:, 0].values
count = 0
with tf.Graph().as_default():
substitute = model.BaselineCNN(hparams, num_classes)
audio_input = tf.placeholder(tf.float32,
shape=[None],
name='audio_input')
start_vars = set(x.name for x in tf.global_variables())
features = inputs.compute_features(audio_input, hparams)
preds = substitute.get_probs(features)
end_vars = tf.global_variables()
model_vars = [x for x in end_vars if x.name not in start_vars]
saver = tf.train.Saver(var_list=model_vars)
with tf.Session() as sess:
saver.restore(sess=sess, save_path=flags.save_model_dir)
for i in range(100):
call = [
'ffmpeg', '-v', 'quiet', '-i',
os.path.join(flags.infer_audio_dir, file_names[i]), '-f',
'f32le', '-ar',
str(44100), '-ac', '1', 'pipe:1'
]
samples = subprocess.check_output(call)
waveform = np.frombuffer(samples, dtype=np.float32)
pr = sess.run([preds], feed_dict={audio_input: waveform})
pr = np.squeeze(pr)
if (pr.ndim != 1):
pr = np.mean(pr, axis=0)
#print(pr)
#print(np.argmax(pr),np.max(pr))
#print(df.iloc[i,1],utils_tf._convert_label_to_label_name(int(np.argmax(pr))))
lab = utils_tf._convert_label_name_to_label(df.iloc[i, 1])
if (lab == np.argmax(pr)):
count += 1
print(lab, np.argmax(pr), np.max(pr))
print(float(count / len(file_names)))
return
def target():
"""
Label the data using a particular model and save the softmax values.
Generates one softmax values per file
"""
flags = parse_flags()
hparams = parse_hparams(flags.hparams)
num_classes = 41
df = pd.read_csv(flags.infer_csv_file)
file_names = df.iloc[:, 0].values
count = 0
sr = 32000
df = pd.read_csv(flags.infer_csv_file)
x, _ = utils_tf._load_dataset(cfg.to_dataset('training'))
generator = utils.fit_scaler(x)
file_names = df.iloc[:, 0].values
with tf.Graph().as_default() as graph:
mel_filt = librosa.filters.mel(sr=32000, n_fft=1024, n_mels=64).T
mel_filt = tf.convert_to_tensor(mel_filt, dtype=tf.float32)
pcm = tf.placeholder(tf.float32, shape=[None], name='input_audio')
model = CleverHansModel(flags.save_model_dir + '.meta', sr, generator,
mel_filt)
saver = model.build_graph(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, flags.save_model_dir)
print(len(file_names))
for i in range(100):
data, _ = utils_tf._preprocess_data(flags.infer_audio_dir,
file_names[i])
l = sess.run([model.get_probs()], feed_dict={pcm: data})
l = np.squeeze(l)
if (l.ndim != 1):
l = np.mean(l, axis=0)
lab = utils_tf._convert_label_name_to_label(df.iloc[i, 1])
if (lab == np.argmax(l)):
count += 1
print(lab, np.argmax(l))
print(count / 100)
def deepfoolcochlear(audio_path,metadata_path,model_path,exp_data_path,adv_audio_path,save_data=False):
#Run the attacks to generate adversarial attacks on manually verified examples on the training and test data
#Load dataset to normalize new data
df = pd.read_csv(metadata_path)
label_names= df.iloc[:,2].values
file_names = df.iloc[:,1].values
sample_rate = 32000
audio_name = []
audio_length = []
original_label = []
original_confidence = []
new_label = []
new_confidence = []
new_o_label_conf = []
snr = []
with tf.Graph().as_default() as graph:
model = CochlearModel(model_path +'.meta')
pcm = tf.placeholder(tf.float32,shape=[None,None],name='input_audio')
saver= model.build_graph(pcm)
deepfool = DFM.DeepFool(model)
deepfool.build_attack(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess,model_path)
for i in range(1,2):
audio_file_name = file_names[i]
try:
data,q = utils_tf._preprocess_data(audio_path,audio_file_name)
data = np.expand_dims(data,axis=0)
except EOFError:
print("EOF Error")
labels= utils_tf._convert_label_name_to_label(label_names[i])
s = sess.run([model.get_probs()],feed_dict={'input_audio:0':data})
s = np.squeeze(s)
if (s.ndim != 1):
s = np.mean(s,axis=0)
print('Ground truth:',labels)
print('Original label number:',np.argmax(s))
print('Original label confidence:',np.max(s))
tic = time.process_time()
adv = deepfool.attack(sess,data,1)
toc = time.process_time()
print('Time for processing sample:',toc-tic,'for iteration:',i)
preds = sess.run([model.get_probs()],feed_dict={pcm:adv})
preds = np.squeeze(preds)
if(preds.ndim !=1):
preds = np.mean(preds,axis=0)
print('New label number:',np.argmax(preds))
print('New label confidence:',np.max(preds))
adv = np.squeeze(adv)
if(save_data):
librosa.output.write_wav(adv_audio_path + 'adv-' + audio_file_name,adv,sample_rate)
audio_name.append(audio_file_name)
audio_length.append(int(q))
original_label.append(np.argmax(s))
original_confidence.append(np.max(s))
new_label.append(np.argmax(preds))
new_confidence.append(np.max(preds))
new_o_label_conf.append(preds[np.argmax(s)])
snr.append(10*np.log10(np.mean(data**2)/(np.mean((adv-data)**2))))
if(save_data):
df_deepfool = pd.DataFrame({'audio_name':audio_name,'audio_length':audio_length,'original_label':original_label,'original_confidence':original_confidence,'new_label':new_label,'new_confidence':new_confidence,'new_orig_conf':new_o_label_conf,'SNR':snr})
with open(exp_data_path,'a') as f:
df_deepfool.to_csv(f,header=False)
def lbfgstargeted(audio_path,
metadata_path,
model_path,
exp_data_path,
adv_audio_path,
save_data=False):
#Run the attacks to generate adversarial attacks on manually verified examples on the training and test data
#Load dataset to normalize new data
x, _ = utils_tf._load_dataset(cfg.to_dataset('training'))
generator = utils.fit_scaler(x)
df = pd.read_csv(metadata_path)
gt_labels = df.iloc[:, 2].values
file_names = df.iloc[:, 1].values
mel_fb = librosa.filters.mel(sr=32000, n_fft=1024, n_mels=64).T
sample_rate = 32000
label_list = [
"Bass_drum", "Cello", "Clarinet", "Oboe", "Snare_drum",
"Violin_or_fiddle"
]
audio_name = []
audio_length = []
original_label = []
original_confidence = []
new_label = []
new_confidence = []
new_o_label_conf = []
snr = []
with tf.Graph().as_default() as graph:
mel_filt = tf.convert_to_tensor(mel_fb, dtype=tf.float32)
model = CleverHansModel(model_path + '.meta', sample_rate, generator,
mel_filt)
pcm = tf.placeholder(tf.float32, shape=[None], name='input_audio')
saver = model.build_graph(pcm)
lbfgs = LB.LBFGS(model, binary_search_steps=2, max_iterations=200)
lbfgs.build_attack(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, model_path)
for i in range(df.shape[0]):
audio_file_name = file_names[i]
try:
data, q = utils_tf._preprocess_data(audio_path,
audio_file_name)
except EOFError:
print("EOF Error")
gt_label = gt_labels[i]
s = sess.run([model.get_probs()],
feed_dict={'input_audio:0': data})
s = np.squeeze(s)
if (s.ndim != 1):
s = np.mean(s, axis=0)
print('Original label number:', np.argmax(s), 'GT:', gt_label)
print('Original label confidence:', np.max(s))
for l in range(len(label_list)):
label = utils_tf._convert_label_name_to_label(label_list[l])
if (label == gt_label):
continue
tic = time.process_time()
adv = lbfgs.attack(sess, data, np.repeat(label, int(q)))
toc = time.process_time()
print('Time for processing sample:', toc - tic,
'for iteration:', i)
preds = sess.run([model.get_probs()], feed_dict={pcm: adv})
preds = np.squeeze(preds)
if (preds.ndim != 1):
preds = np.mean(preds, axis=0)
print('New label number:', np.argmax(preds))
print('New label confidence:', np.max(preds))
if (save_data):
librosa.output.write_wav(
adv_audio_path + 'adv-' + label_list[l] + '-' +
audio_file_name, adv, sample_rate)
audio_name.append(audio_file_name)
audio_length.append(int(q))
original_label.append(np.argmax(s))
original_confidence.append(np.max(s))
new_label.append(np.argmax(preds))
new_confidence.append(np.max(preds))
new_o_label_conf.append(preds[np.argmax(s)])
snr.append(
10 *
np.log10(np.mean(data**2) / (np.mean((adv - data)**2))))
if (save_data):
df_deepfool = pd.DataFrame({
'audio_name': audio_name,
'audio_length': audio_length,
'original_label': original_label,
'original_confidence': original_confidence,
'new_label': new_label,
'new_confidence': new_confidence,
'new_orig_conf': new_o_label_conf,
'SNR': snr
})
with open(exp_data_path, 'a') as f:
df_deepfool.to_csv(f, header=False)
def inferencecochlear(audio_path,
metadata_path,
model_path,
exp_data_path,
adv_audio_path,
save_data=False):
#Run the attacks to generate adversarial attacks on manually verified examples on the training and test data
#Load dataset to normalize new data
df = pd.read_csv(metadata_path)
label_names = df.iloc[:, 1].values
file_names = df.iloc[:, 0].values
sample_rate = 32000
audio_name = []
inferred_label = []
inferred_confidence = []
ground_truth = []
with tf.Graph().as_default() as graph:
model = CochlearModel(model_path + '.meta')
pcm = tf.placeholder(tf.float32,
shape=[None, None],
name='input_audio')
saver = model.build_graph(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, model_path)
count = 0
count_tot = 0
for i in range(df.shape[0]):
audio_file_name = file_names[i]
try:
data, q = utils_tf._preprocess_data(audio_path,
audio_file_name)
data = np.expand_dims(data, axis=0)
except EOFError:
print("EOF Error")
gt_label = utils_tf._convert_label_name_to_label(label_names[i])
s = sess.run([model.get_probs()],
feed_dict={'input_audio:0': data})
s = np.squeeze(s)
if (s.ndim != 1):
s = np.mean(s, axis=0)
count_tot += 1
if (gt_label == np.argmax(s)):
count += 1
if (i % 1000 == 0):
print('Iteration number:', i)
print('Current accuracy:', float(count / count_tot))
audio_name.append(audio_file_name)
inferred_label.append(np.argmax(s))
inferred_confidence.append(np.max(s))
ground_truth.append(gt_label)
if (save_data):
df_infer = pd.DataFrame({
'audio_name': audio_name,
'ground_truth': ground_truth,
'inferred_label': inferred_label,
'inferred_confidence': inferred_confidence
})
with open(exp_data_path, 'w') as f:
df_infer.to_csv(f, header=False)
def inferenceiqbal(audio_path,
metadata_path,
model_path,
exp_data_path,
adv_audio_path,
save_data=False):
#Run the attacks to generate adversarial attacks on manually verified examples on the training and test data
#Load dataset to normalize new data
x, _ = utils_tf._load_dataset(cfg.to_dataset('training'))
generator = utils.fit_scaler(x)
df = pd.read_csv(metadata_path)
label_names = df.iloc[:, 1].values
file_names = df.iloc[:, 0].values
mel_fb = librosa.filters.mel(sr=32000, n_fft=1024, n_mels=64).T
sample_rate = 32000
audio_name = []
ground_truth = []
inferred_label = []
inferred_confidence = []
with tf.Graph().as_default() as graph:
mel_filt = tf.convert_to_tensor(mel_fb, dtype=tf.float32)
model = CleverHansModel(model_path + '.meta', sample_rate, generator,
mel_filt)
pcm = tf.placeholder(tf.float32, shape=[None], name='input_audio')
saver = model.build_graph(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, model_path)
count = 0
count_tot = 0
for i in range(df.shape[0]):
audio_file_name = file_names[i]
try:
data, q = utils_tf._preprocess_data(audio_path,
audio_file_name)
except EOFError:
print("EOF Error")
gt_label = utils_tf._convert_label_name_to_label(label_names[i])
s = sess.run([model.get_probs()],
feed_dict={'input_audio:0': data})
s = np.squeeze(s)
if (s.ndim != 1):
s = np.mean(s, axis=0)
label = np.argmax(s)
count_tot += 1
if (label == gt_label):
count += 1
if (i % 1000 == 0):
print('Iteration number:', i)
print('Current accuracy:', float(count / count_tot))
audio_name.append(audio_file_name)
ground_truth.append(gt_label)
inferred_label.append(label)
inferred_confidence.append(np.max(s))
if (save_data):
df_deepfool = pd.DataFrame({
'audio_name':
audio_name,
'ground_truth':
ground_truth,
'inferred_label':
inferred_label,
'inferred_confidence':
inferred_confidence
})
with open(exp_data_path, 'w') as f:
df_deepfool.to_csv(f, header=False)
def carliniwagneruntargetedcochlear(audio_path,
metadata_path,
model_path,
exp_data_path,
adv_audio_path,
save_data=False):
#Run the attacks to generate adversarial attacks on manually verified examples on the training and test data
#Load dataset to normalize new data
df = pd.read_csv(metadata_path)
label_names = df.iloc[:, 2].values
file_names = df.iloc[:, 1].values
sample_rate = 32000
audio_name = []
audio_length = []
original_label = []
original_confidence = []
new_label = []
new_confidence = []
new_o_label_conf = []
snr = []
with tf.Graph().as_default() as graph:
model = CochlearModel(model_path + '.meta')
pcm = tf.placeholder(tf.float32,
shape=[None, None],
name='input_audio')
carliniwagner = CW.CarliniWagnerAttack(model,
learning_rate=1e-5,
initial_const=1e-2,
max_iterations=1000,
confidence=500,
binary_search_steps=2)
saver = carliniwagner.build_attack(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, model_path)
for i in range(df.shape[0]):
audio_file_name = file_names[i]
try:
data, q = utils_tf._preprocess_data(audio_path,
audio_file_name)
data = np.expand_dims(data, axis=0)
except EOFError:
print("EOF Error")
label = utils_tf._convert_label_name_to_label(label_names[i])
print('Ground truth label:', label_names[i], label)
tic = time.process_time()
adv, o_label, o_conf, n_label, n_conf, n_conf_gt = carliniwagner.attack(
sess, data, label, label, 1, prob_thresh=0.0244)
toc = time.process_time()
print('Time for iteration:', i, 'is', toc - tic)
adv = np.squeeze(adv)
if (save_data):
librosa.output.write_wav(
adv_audio_path + 'adv-' + audio_file_name, adv,
sample_rate)
audio_name.append(audio_file_name)
audio_length.append(int(q))
original_label.append(o_label)
original_confidence.append(o_conf)
new_label.append(n_label)
new_confidence.append(n_conf)
new_o_label_conf.append(n_conf_gt)
snr.append(10 *
np.log10(np.mean(data**2) / (np.mean((adv - data)**2))))
if (save_data):
df_cw = pd.DataFrame({
'audio_name': audio_name,
'audio_length': audio_length,
'original_label': original_label,
'original_confidence': original_confidence,
'new_label': new_label,
'new_confidence': new_confidence,
'new_orig_conf': new_o_label_conf,
'SNR': snr
})
with open(exp_data_path, 'w') as f:
df_cw.to_csv(f, header=False)
def carliniwagnertargeted(audio_path,
metadata_path,
model_path,
exp_data_path,
adv_audio_path,
save_data=False):
#Run the attacks to generate adversarial attacks on manually verified examples on the training and test data
#Load dataset to normalize new data
x, _ = utils_tf._load_dataset(cfg.to_dataset('training'))
generator = utils.fit_scaler(x)
df = pd.read_csv(metadata_path)
gt_labels = df.iloc[:, 2].values
file_names = df.iloc[:, 1].values
mel_fb = librosa.filters.mel(sr=32000, n_fft=1024, n_mels=64).T
sample_rate = 32000
label_list = [
"Bass_drum", "Cello", "Clarinet", "Oboe", "Snare_drum",
"Violin_or_fiddle"
]
audio_name = []
audio_length = []
original_label = []
original_confidence = []
new_label = []
new_confidence = []
new_o_label_conf = []
snr = []
with tf.Graph().as_default() as graph:
mel_filt = tf.convert_to_tensor(mel_fb, dtype=tf.float32)
model = CleverHansModel(model_path + '.meta', sample_rate, generator,
mel_filt)
pcm = tf.placeholder(tf.float32, shape=[None], name='input_audio')
carliniwagner = CW.CarliniWagnerAttack(model,
learning_rate=1e-5,
confidence=500,
targeted=True,
max_iterations=1000,
binary_search_steps=2)
saver = carliniwagner.build_attack(pcm)
with tf.Session(graph=graph) as sess:
saver.restore(sess, model_path)
for i in range(df.shape[0]):
audio_file_name = file_names[i]
try:
data, q = utils_tf._preprocess_data(audio_path,
audio_file_name)
except EOFError:
print("EOF Error")
gt_label = gt_labels[i]
print('Ground truth label:', gt_label, 'Audio_file:',
file_names[i])
for l in range(len(label_list)):
label = utils_tf._convert_label_name_to_label(label_list[l])
if (label == gt_label):
continue
adv, o_label, o_conf, n_label, n_conf, n_gt_conf = carliniwagner.attack(
sess,
data,
label,
np.repeat(label, int(q)),
int(q),
prob_thresh=0.975)
if (save_data):
librosa.output.write_wav(
adv_audio_path + 'adv-' + label_list[l] + '-' +
audio_file_name, adv, sample_rate)
audio_name.append(audio_file_name)
audio_length.append(int(q))
original_label.append(o_label)
original_confidence.append(o_conf)
new_label.append(n_label)
new_confidence.append(n_conf)
new_o_label_conf.append(n_gt_conf)
snr.append(
10 * np.log10(np.mean(data**2) / (np.mean(adv - data)**2)))
if (save_data):
df_cw = pd.DataFrame(
{
'audio_name': audio_name,
'audio_length': audio_length,
'original_label': original_label,
'original_confidence': original_confidence,
'new_label': new_label,
'new_confidence': new_confidence,
'SNR': snr
},
columns=[
'audio_name', 'audio_length', 'original_label',
'original_confidence', 'new_label', 'new_confidence', 'SNR'
])
with open(exp_data_path, 'a') as f:
df_cw.to_csv(f)
def main():
flags = parse_flags()
hparams = parse_hparams(flags.hparams)
print(hparams)
num_classes=41
df = pd.read_csv(flags.infer_csv_file)
file_names = df.iloc[:,0].values
labels = df.iloc[:,1].values
gt = []
orig_label = []
adv_label = []
orig_pred = []
adv_pred = []
snr = []
files = []
t = []
with tf.Graph().as_default() and tf.Session() as sess:
if(hparams.vgg13_features):
substitute_model = model.vgg13(hparams,num_classes)
else:
substitute_model = model.BaselineCNN(hparams,num_classes)
cw = CW.CarliniWagnerAttack(model=substitute_model,save_model_dir=flags.save_model_dir,sess=sess,hparams=hparams)
cw.build_attack()
for i in range(len(file_names)):
start_time=time.time()
#call = ['ffmpeg','-v','quiet','-i',os.path.join(flags.infer_audio_dir,file_names[i]),'-f','f32le', '-ar',str(hparams.sample_rate),'-ac','1','pipe:1']
#samples = subprocess.check_output(call)
#data = np.frombuffer(samples, dtype=np.float32)
data,_ = utils_tf._preprocess_data(flags.infer_audio_dir,file_names[i])
lab = utils_tf._convert_label_name_to_label(labels[i])
print(data.shape)
if(hparams.targeted):
set_target=False
while(not set_target):
target_label = np.random.randint(41)
if(lab != target_label):
set_target=True
else:
target_label = lab
audio,pred,pred_orig,noise = cw.attack(data,target_label)
if(audio is None):
continue
print('TIME IN SECONDS!',time.time()-start_time)
gt.append(lab)
orig_label.append(np.argmax(pred_orig))
orig_pred.append(np.max(pred_orig))
adv_label.append(np.argmax(pred))
adv_pred.append(np.max(pred))
snr.append(noise)
files.append(file_names[i])
t.append(time.time()-start_time)
wav.write(os.path.join(flags.write_audio_dir,file_names[i]),44100,audio)
df_out = pd.DataFrame({'fname':files,'gt':gt,'original_label':orig_label,'original_pred':
orig_pred,'adv_label':adv_label,'adv_pred':adv_pred,'snr':snr,'time':t})
df_out.to_csv('adv_data_vgg13.csv',index=False)
