attack result
原始精确度 Test net output #0: accuracy = 0.9658 #Test net output #0: accuracy = 0.9607
原始weight加 eps =5的噪音，精确度到 error=0.1755 I0725 20:58:48.742980 6007 caffe.cpp:318] Loss: 0.430706 I0725 20:58:48.742991 6007 caffe.cpp:330] accuracy = 0.854 I0725 20:58:48.743017 6007 caffe.cpp:330] loss = 0.430706 (* 1 = 0.430706 loss)
生成weight为ip1_t.npy, ip-2_t.npy， attack结果为不可看。
!!!! ———————————— 在attack的程序中需要保存这段，不然是从0.1 0.1 0.1 开始 for _ in range(1000): batch_xs, batch_ys = mnist.train.next_batch(100) [tmp_train,tmp_w1,tmp_w2] = sess.run([train_step,w1,w2], feed_dict={x: batch_xs, y_: batch_ys})#这里的tmp_w1,tmp_w2是训练好的weight