Adversarial attack on Detector

[AlexeyAB]

Try to find Dog, Car, Bicycle on this image: drawn.zip

By using Yolov3-model ./darknet detector test cfg/coco.data yolov3.cfg yolov3.weights drawn.png

---

How to use:

1. Build Darknet with GPU=1 CUDNN=1 OPENCV=1

2. Run command: ./darknet detector draw cfg/coco.data yolov3.cfg yolov3.weights -thresh 0.25 dog.jpg

3. Move trackbar:

   • iterations=200
   • learning_rate exp = 10
   • class_id = 15

4. Select any area on the image by Left-mouse-button

5. Will be saved image drawn.png

[sctrueew]

@AlexeyAB Hi,

I could detect the dog with these parameters. Can you explain more about this feature?

[AlexeyAB]

@zpmmehrdad

Can you explain more about this feature?

• Regular training - network during training changes their weights to required detect objects
• Adversarial attack - network during training changes initial image to required detect objects, so you can make this neural network detect the objects you need, and not detect unnecessary

I improved adversarial attack and training. Download the latest Darknet code. Now use default settings.

[HagegeR]

this is a kind of transfer learning or fine tuning? won't it affect global performance (improving one class while degrading others) ?

[AlexeyAB]

@HagegeR This is a utility for this to fool the detector. It does't train the model, it train the image. To train the model use this: https://github.com/AlexeyAB/darknet/issues/5117

[sisrfeng]

@zpmmehrdad

Can you explain more about this feature?

• Regular training - network during training changes their weights to required detect objects
• Adversarial attack - network during training changes initial image to required detect objects, so you can make this neural network detect the objects you need, and not detect unnecessary

I improved adversarial attack and training. Download the latest Darknet code. Now use default settings.

Should "to required detect objects" be "to detect required objects" ? By adding some noise on the dog to make it look like a cat (to machine but not to human), then training the model to detect the modified dog , "so you can make this neural network detect the objects you need (dog)" , right? Similar to this?

[AlexeyAB]

@sisrfeng Yes, but since we have access to the structure+weights of neural network (not just black-box), so we can do such attack more efficient (with less changes - like side effects of image-compression). But we can also teach it to defend herself against such attacks.