darknet

darknet

FGSM

Fast Gradient Sign Attatck(FGSM) 出自 Explaining and Harnessing Adversarial Examples.沿梯度正方向修改输入数据，达到迷惑模型的目的。

流程说明

• 计算网络error相对于输入数据的梯度
  forward($data_{old}$)
  grad = backword()
• 沿梯度正方向修改输入数据
  $data_{new} = data_{old} + eps \times sign(grad)$
• 计算修改后数据的预测结果
  forward($data_{new}$)

结果

随着eps的增加，模型在新数据上预测准确率逐渐降低

使用方式

./darknet classifier fgsm cfg/mnist.data cfg/mnist.cfg backup/mnist_5.weights -eps 0.1

yolov3训练

修改example代码 train.sh/test.sh

基于voc数据集格式+预训练模型

./darknet detector train examples/traffic/image.data examples/traffic/yolov3.cfg pretrained/darknet53.conv.74 2>&1 | tee examples/traffic/train.log

统计模型recalling

./darknet detector recall2 examples/traffic/image.data examples/traffic/yolov3.cfg examples/traffic/yolov3_900.weights examples/traffic/test.txt 2>&1 | tee examples/traffic/yolov3_900.test.log

批量运行，保存结果

./darknet detector test2 examples/traffic/image.data examples/traffic/yolov3.cfg examples/traffic/yolov3_800.weights examples/traffic/test.txt -out predict/