The security problem is that when we transmit those intermediate features to the cloud server, we need to consider the man-in-the-middle attack, who can tap into the features and use Generative Adversarial Networks to restore the private training data from the features. The possibility that the attacker can recover the original data declines monotonically regarding the depth of the neural network. For VGG16, the possibility is conv3-64, conv3-64 (90%); conv3-128, conv3-128(70%); conv3-256, conv3-256, conv3-256 (40%); conv3-512, conv3-512, conv3-512 (20%); conv3-512, conv3-512, conv3-512 (5%), for each filter module, the number indicates the number of filters (e.g., 512 filters) and the percentage indicates the attack success rate (
The security problem is that when we transmit those intermediate features to the cloud server, we need to
consider the man-in-the-middle attack, who can tap into the features and use Generative Adversarial
Networks to restore the private training data from the features. The possibility that the attacker can
recover the original data declines monotonically regarding the depth of the neural network. For VGG16,
the possibility is conv3-64, conv3-64 (90%); conv3-128, conv3-128(70%); conv3-256, conv3-256,
conv3-256 (40%); conv3-512, conv3-512, conv3-512 (20%); conv3-512, conv3-512, conv3-512 (5%), for
each filter module, the number indicates the number of filters (e.g., 512 filters) and the percentage
indicates the attack success rate (20%).
Now we want to achieve a balance between security and network bandwidth. For example, if we are
transferring the first layer of 224x224x3, we are transferring 220x220x3x64 pixels with (1-90%) in the
security gauge. We use a simple weighted formular to weigh the network bandwidth and security, Y = N
+ a* S, where N = 220x220x3x64 + a*0.1, set a = 10^6.
Compute the Y value for each filter bank. Note that you will need to calculate the correct output size from
the convolutional layer. For example, the output size from the first conv3 layer is 224-3+1 = 222. After
the second conv3 layer is 220x220x3. Show the computation for all the 64,128,256 and 512 layer
modules (assume we partition the network at the end of each module).
Trending now
This is a popular solution!
Step by step
Solved in 2 steps with 1 images
