Protection of neural networks by obfuscation of activation functions

Patent metadata

Field	Value
Publication number	US-12099622-B2
Application number	US-202117553536-A
Country	US
Kind code	B2
Filing date	Dec 16, 2021
Priority date	Dec 21, 2020
Publication date	Sep 24, 2024
Grant date	Sep 24, 2024

Abstract

Aspects of the present disclosure involve implementations that may be used to protect neural network models against adversarial attacks by obfuscating neural network operations and architecture. Obfuscation techniques include obfuscating weights and biases of neural network nodes, obfuscating activation functions used by neural networks, as well as obfuscating neural network architecture by introducing dummy operations, dummy nodes, and dummy layers into the neural networks.

First claim

What is claimed is:

1. A method to execute a neural network model that includes a plurality of nodes, the method comprising:
determining, by a processing device, based on parameters of a first node of the plurality of nodes, a weighted input into an activation function for the first node;
selecting, by the processing device, an obfuscation function for the first node;
determining, by the processing device, a first composite activation function for the first node, wherein the first composite activation function is formed by the activation function for the first node and the obfuscation function for the first node;
applying, by the processing device, the first composite activation function to the weighted input to compute an obfuscated output of the first node; and
providing, by the processing device, the obfuscated output of the first node to a second node of the plurality of nodes of the neural network.

2. The method of claim 1 , wherein the obfuscation function is an invertible function.

3. The method of claim 1 , wherein the weighted input into the activation function of the first node is obtained using a plurality of masked weights of the first node.

4. The method of claim 1 , further comprising: determining a weighted input into an activation function of the second node, by applying, to the provided obfuscated output of the first node, a weight of the second node composite with a de-obfuscation function.

5. The method of claim 4 , wherein the weighted input into the activation function of the second node is obtained using a plurality of masked weights of the second node.

6. The method of claim 1 , wherein the first composite activation function is one of a plurality of composite activation functions for the first node, each of the plurality of composite activation functions is based on a respective activation function of a plurality of activation functions for the first node and a respective obfuscation function of a plurality of obfuscation functions for the first node, the method further comprising:
applying, by the processing device, each of the plurality of composite activation functions to the weighted input to compute a respective obfuscated output of a plurality of obfuscated outputs of the first node; and
providing, by the processing device, each of the plurality of obfuscated outputs to the second node.

7. The method of claim 6 , further comprising:
prior to providing each of the plurality of obfuscated outputs to the second node, masking, by the processing device, the plurality of obfuscated outputs.

8. The method of claim 7 , further comprising:
unmasking, by the processing device, the masked plurality of obfuscated outputs, wherein the unmasking is composite with determining a weighted input into the second node.

9. The method of claim 8 , wherein the unmasking is further composite with one or more de-obfuscation functions.

10. A method comprising:
identifying a neural network (NN) model to be protected against adversarial attacks, wherein the NN model includes a plurality of nodes; and
modifying the NN model to obtain a modified NN model, wherein each of the NN model and the modified NN model is configured to output same target output based on same input, wherein the modified NN model comprises:
determining, by a processing device, based on parameters of a first node of the plurality of nodes, a weighted input into an activation function for the first node;
selecting, by the processing device, an obfuscation function for the first node;
determining, by the processing device, a first composite activation function for the first node, wherein the first composite activation function is formed by the activation function for the first node and the obfuscation function for the first node;
applying, by the processing device, the first composite activation function to the weighted input to compute an obfuscated output of the first node; and
providing, by the processing device, the obfuscated output of the first node to a second node of the plurality of nodes of the neural network.

11. The method of claim 10 , wherein the weighted input into the activation function of the first node is obtained using a plurality of masked weights of the first node.

12. A system to execute a neural network model that includes a plurality of nodes, the system comprising:
a memory device; and
a processing device communicatively coupled to the memory device, the processing device to:
determine, based on parameters of a first node of the plurality of nodes, a weighted input into an activation function for the first node;
select an obfuscation function for the first node;
determine a first composite activation function for the first node, wherein the first composite activation function is formed by the activation function for the first node and the obfuscation function for the first node;
apply the first composite activation function to the weighted input to compute an obfuscated output of the first node; and
provide the obfuscated output of the first node to a second node of the plurality of nodes of the neural network.

13. The system of claim 12 , wherein the obfuscation function is an invertible function.

14. The system of claim 12 , wherein the weighted input into the activation function of the first node is obtained using a plurality of masked weights of the first node.

15. The system of claim 12 , wherein the processing device is further to: determine a weighted input into an activation function of the second node, by applying, to the provided obfuscated output of the first node, a weight of the second node composite with a de-obfuscation function.

16. The system of claim 15 , wherein the weighted input into the activation function of the second node is obtained using a plurality of masked weights of the second node.

17. The system of claim 12 , wherein the first composite activation function is one of a plurality of composite activation functions for the first node, each of the plurality of composite activation functions is based on a respective activation function of a plurality of activation functions for the first node and a respective obfuscation function of a plurality of obfuscation functions for the first node, and wherein the processing device is further to:
apply each of the plurality of composite activation functions to the weighted input to compute a respective obfuscated output of a plurality of obfuscated outputs of the first node; and
provide each of the plurality of obfuscated outputs to the second node.

18. The system of claim 17 , wherein the processing device is further to: mask the plurality of obfuscated outputs prior to providing each of the plurality of obfuscated outputs to the second node.

19. The system of claim 18 , wherein the processing device is further to: perform unmasking of the masked plurality of obfuscated outputs, wherein the unmasking is composite with determining a weighted input into the second node.

20. The system of claim 19 , wherein the unmasking is further composite with one or more de-obfuscation functions.

Assignees

• Cryptography Res Inc

Inventors

• Marson Mark Evan
• Hamburg Michael Alexander
• Handschuh Helena

Classifications

• G06V10/82

  using neural networks · CPC title

• G06F21/14

  against software analysis or reverse engineering, e.g. by obfuscation · CPC title

• G06F21/6227

  where protection concerns the structure of data, e.g. records, types, queries · CPC title

• G06N3/02

  Neural networks · CPC title

• G06N3/048

  Activation functions · CPC title