def __init__(
self,
o_audio_file,
t_audio_file,
save_p_fname,
save_info_fname,
MODE="TIMIE",
dct_field=0.65,
):
self.o_audio_file = o_audio_file
self.t_audio_file = t_audio_file
self.model = Sincnet.get_speaker_model(MODE)
self.model = self.model.eval()
self.speaker_label, self.label_speaker = Sincnet.get_speaker_label(
MODE)
self.save_p_fname = save_p_fname
self.save_info_fname = save_info_fname
self.dct = lambda x: ffp.dct(x, norm='ortho')
self.idct = lambda ix: ffp.idct(ix, norm='ortho')
self.o_audio, self.sr = sf.read(o_audio_file)
self.t_audio, self.sr = sf.read(t_audio_file)
# 目标label以及原始label以及name
self.o_label = self.predict_one_label(self.o_audio)
self.t_label = self.predict_one_label(self.t_audio)
self.o_name = self.get_name_by_label(self.o_label)
self.t_name = self.get_name_by_label(self.t_label)
# 初始化音频
self.audio_len = min(len(self.o_audio), len(self.t_audio))
self.o_audio = self.o_audio[:self.audio_len]
self.t_audio = self.t_audio[:self.audio_len]
self.o_audio /= np.linalg.norm(self.o_audio, np.inf)
self.t_audio /= np.linalg.norm(self.t_audio, np.inf)
# 初始化音频微缩
self.o_audio *= 0.95
self.t_audio *= 0.95
self.o2_audio = None
# 定义扰动部分变量
self.best_pretub_scale = float('inf')
self.best_pretub = None
self.best_clip_scale = float('inf')
self.best_clip_perturb = None
self.interval = [None, None]
self.clip_perturb_len = None
# 超参
self.query_num = 0
self.theta = 1e-4 # 二分结束条件
self.dct_field = dct_field
def __init__(self,
MODE,
SAVE_DIR_PATH,
query_num,
interval,
num_iterations=2000,
gamma=1.0,
stepsize_search='geometric_progression',
max_num_evals=1e4,
init_num_evals=100,
query_limit=25000,
verbose=True):
self.model = Sincnet.get_speaker_model(MODE)
self.num_iterations = num_iterations
self.gamma = gamma
self.stepsize_search = stepsize_search
self.max_num_evals = max_num_evals
self.init_num_evals = init_num_evals
self.verbose = verbose
self.SAVE_DIR_PATH = SAVE_DIR_PATH
self.query_num = query_num
self.query_limit = query_limit
self.interval = interval
self.dct = lambda x: ffp.dct(x, norm='ortho')
self.idct = lambda ix: ffp.idct(ix, norm='ortho')
def predict_one_label(self, data):
self.query_num += 1
data = data.squeeze()
pred_real, pred_pro = Sincnet.sentence_test(
self.model,
torch.from_numpy(data).float().cuda())
return pred_real
def __init__(self,
MODE,
SAVE_DIR_PATH,
num_iterations=2000,
gamma=1.0,
stepsize_search='geometric_progression',
max_num_evals=1e4,
init_num_evals=100,
query_limit=25000,
verbose=True):
self.model = Sincnet.get_speaker_model(MODE)
self.num_iterations = num_iterations
self.gamma = gamma
self.stepsize_search = stepsize_search
self.max_num_evals = max_num_evals
self.init_num_evals = init_num_evals
self.verbose = verbose
self.SAVE_DIR_PATH = SAVE_DIR_PATH
self.query_num = 0
self.query_limit = query_limit
# -- coding: utf-8 --
from utils import Sincnet
import torch
import numpy as np
import soundfile as sf
import librosa.display
import matplotlib.pyplot as plt
import pickle
import utils
import os
import LOCAL_ATT_HSJA_ATTACK
MODE = "TIMIT"
abs_path=os.getcwd()
model = Sincnet.get_speaker_model(MODE)
speaker_label, label_speaker = Sincnet.get_speaker_label(MODE)
def mkd(name):
if os.path.exist(name)==False:
os.mkdir(os.path.join(abs_path,name))
if MODE=="Librispeech":
save_dir = "lahresult\lib"
save_adv_dir="lahresult\libaudio"
mkd(save_dir)
mkd(save_adv_dir)
attackdir = r"AttackDataset\lib-attack-audio"
targetdir = r"AttackDataset\lib-target-audio"
else:
save_dir = r"lahresult\timit"
save_adv_dir = r"lahresult\timitaudio"
mkd(save_dir)
def show_max(wav):
qq,_ = Sincnet.sentence_test_lib(model,wav.float().cuda())
print(qq)
# -*- coding: UTF-8 -*-
from utils import Sincnet
#from utils.sincet_for_lib import *
import torch
import numpy as np
import soundfile as sf
import os
import time
from LocalSearch.select_region_attack_timit import select_region
MODE = "Librispeech"
#model = get_speaker_model2(MODE)
model = Sincnet.get_speaker_model(MODE)
speaker_label, label_speaker = Sincnet.get_speaker_label(MODE)
def show_max(wav):
qq,_ = Sincnet.sentence_test_lib(model,wav.float().cuda())
print(qq)
adir=r"F:\SR-ATK\用于画图sampleaudio\lib-attack-audio"
tdir=r"F:\SR-ATK\用于画图sampleaudio\lib-target-audio"
o_audio_files = os.listdir(adir)
o_audio_files.sort(key=lambda item: int(item.split('-')[0]))
t_audio_files = os.listdir(tdir)
t_audio_files.sort(key=lambda item: int(item.split('-')[0]))
save_txt_path=r"F:\SR-ATK\占比记录\newlib.txt"
record=[]#[非目标分类区占比 ， 攻击区域占比，原初始扰动大小，更新的初始扰动大小]
for a,t in zip(o_audio_files,t_audio_files):
a=os.path.join(adir,a)
t=os.path.join(tdir,t)
def predict_one_label(model, audio):
qq, _ = Sincnet.sentence_test_lib(model, audio.float().cuda())
return qq
def __init__(self, MODE, savepath, limit_count=25000):
self.querynum = 0
self.model = Sincnet.get_speaker_model(MODE)
self.savepath = savepath
self.limit_count = limit_count
def predict_label(self,data):
pred_real, pred_pro = Sincnet.sentence_test(self.model,
torch.from_numpy(data).float().cuda())
return pred_real