can a convolutional neural network be reverse engineered?

Question

can a modern,complex convolutional neural network for object detection be reverse engineered to acquire the original image training data? Thanks.

i did try searching for my question online but the answers didn't exactly answer my question

Answer 1

Yes,of course. When you add different filteration techniques like average pooling, max pooling, the spatial dimension of your image will obviously decrease until the Same padding is set. This techniques is known as downsampling. So to aquire or preserve the original size, there are different Upsampling techniques. These are some common Upsampling techniques:

1. Nearest neighbor: It is a simple upsampling techniques. Related or common data in rows and column are copied near to them until original size meets.

2. Unpooling: This techniques reverse the every techniques that was used during downsampling. It is unpooling the pooling layer that we have applied before. It would not guarantee exact same pixel but approximate resolutin can be obtain.

3. Transpose convolution: So you know, convolution layer was multiplied with pooling layer and the image lost its dimension. What Transpose convolution does is it transpose what happen during regular convolution operation also known as deconvolution. It performs the opposite operation of regular convolution, allowing the network to learn to increase the resolution of the feature maps.

If you required additional information I recommend you to go through it. Downsampling and Upsampling in CNN