Secure convolutional neural networks (CNN) accelerator

Patent metadata

Field	Value
Publication number	US-10706349-B2
Application number	US-201715730316-A
Country	US
Kind code	B2
Filing date	Oct 11, 2017
Priority date	May 25, 2017
Publication date	Jul 7, 2020
Grant date	Jul 7, 2020

Abstract

A Convolutional Neural Network (CNN) based-signal processing includes receiving of an encrypted output from a first layer of a multi-layer CNN data. The received encrypted output is subsequently decrypted to form a decrypted input to a second layer of the multi-layer CNN data. A convolution of the decrypted input with a corresponding decrypted weight may generate a second layer output, which may be encrypted and used as an encrypted input to a third layer of the multi-layer CNN data.

First claim

What is claimed is:

1. A method of signal processing comprising:
receiving an encrypted output from a first layer of a multi-layer convolutional neural networks (CNN) data;
decrypting the received encrypted output to form a decrypted input to a second layer of the multi-layer CNN data using a first cryptographic key, the first cryptographic key is part of a first cryptographic key pair;
receiving an encrypted weight that corresponds to the decrypted input;
decrypting the received encrypted weight uses a second cryptographic key, the second cryptographic key is part of a second cryptographic key pair, and the second cryptographic key pair is different than the first cryptographic key pair;
performing a convolution on the decrypted input and the decrypted weight to generate a second layer output;
and encrypting the second layer output to form an encrypted second layer output, wherein the encrypted second layer output is used as an encrypted input to a third layer of the multi-layer CNN data.

2. The method of claim 1 , wherein the second layer is any layer of the multi-layer CNN data.

3. The method of claim 1 , wherein the encrypting the second layer output is not performed if the second layer is a last layer of the multi-layer CNN data.

4. The method of claim 1 , wherein the decrypting the received encrypted output and the encrypting are symmetric key decryption and encryption, respectively.

5. The method of claim 1 , wherein the decrypting the received encrypted output and the encrypting are on-the-fly decryption and encryption, respectively.

6. The method of claim 1 , wherein the first and second cryptographic keys are different and the first and second cryptographic keys are stored at an external memory using an asymmetrical encryption.

7. The method of claim 1 , wherein different cryptographic keys are used for decrypting different layers of the multi-layer CNN data, and wherein different cryptographic keys are used for encrypting different layers of the multi-layer CNN data.

8. The method of claim 1 , wherein the encrypted weight is stored at an external memory.

9. The method of claim 1 , the decrypting of the encrypted weight utilizes a block-chaining to provide higher security.

10. The method of claim 1 , wherein an input to the first layer is an unencrypted layer of an image frame.

11. A silicon-on-chip (SoC) device comprising;
an input feature decryption block configured to:
receive an encrypted output from a first layer of a multi-layer convolutional neural networks (CNN) data; and
decrypt the received encrypted output using a first cryptographic key, wherein the decrypted output is used as an input of a second layer of the multi-layer CNN data and the first cryptographic key is part of a first cryptographic key pair;
a weight kernel decryption block configured to:
receive an encrypted weight that corresponds to the input of the second layer;

and decrypt the received encrypted weight using a second cryptographic key, the second cryptographic key is part of a second cryptographic key pair, and the second cryptographic key pair is different than the first cryptographic key pair;
a CNN hardware (HW) core configured to perform a convolution of the input and the decrypted weight to generate a second layer output;
and an output feature encryption block configured to encrypt the second layer output to form an encrypted second layer output, wherein the encrypted second layer output is used as an encrypted input to a third layer of the multi-layer CNN data.

12. The SoC device of claim 11 further comprising a key management that is configured to provide the first cryptographic key to the input feature decryption block.

13. The SoC device of claim 12 , wherein the first cryptographic key is stored using an asymmetrical encryption in a memory that is external to the SoC device.

14. The SoC device of claim 11 , wherein the encrypted weight is stored in a memory that is external to the SoC device.

15. The SoC device of claim 11 , the decrypting of the encrypted weight utilizes a block-chaining to provide higher security.

16. The SoC device of claim 15 , wherein decrypting of the encrypted output is performed by disabling the block-chaining to allow easy random access.

17. A device comprising;
an external memory; and
a silicon-on-chip (SoC) device coupled to the external memory, the SoC device comprising:
an input feature decryption block configured to:
receive an encrypted output from a first layer of a multi-layer convolutional neural networks (CNN) data; and
decrypt the received encrypted output using a first cryptographic key, wherein the decrypted output is used as an input of a second layer of the multi-layer CNN data and the first cryptographic key is part of a first cryptographic key pair;
a weight kernel decryption block configured to:
receive an encrypted weight from the external memory, the received encrypted weight corresponding to the input of the second layer; and
decrypt the received encrypted weight using a second cryptographic key, the second cryptographic key is part of a second cryptographic key pair, and the second cryptographic key pair is different than the first cryptographic key pair;
a CNN hardware (HW) core configured to perform a convolution of the input and the decrypted weight to generate a second layer output;
an output feature encryption block configured to encrypt the second layer output to form an encrypted second layer output, wherein the encrypted second layer output is used as an encrypted input to a third layer of the multi-layer CNN data.

18. The device of claim 17 , wherein the first cryptographic key is stored using an asymmetrical encryption in the external memory.

19. The device of claim 17 , wherein the first cryptographic key is a master key used for encryption and decryption and stored within the SoC device and not readable by software.

20. The device of claim 17 , further comprising a key management memory configured to store the first cryptographic key and the second cryptographic key.

21. The device of claim 17 , wherein the decrypting of the encrypted weight utilizes a block-chaining to provide higher security.

Assignees

• Texas Instruments Inc

Inventors

• Mody Mihir Narendra
• Raju Veeramanikandan
• Ghone Chaitanya
• Poddar Deepak

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464 Primary

  Convolutional networks [CNN, ConvNet] · CPC title

• G06F21/602

  Providing cryptographic facilities or services · CPC title

• H04L63/0435 Primary

  wherein the sending and receiving network entities apply symmetric encryption, i.e. same key used for encryption and decryption (cryptographic mechanisms or cryptographic arrangements for symmetric key encryption H04L9/06) · CPC title

• H04L63/0464

  using hop-by-hop encryption, i.e. wherein an intermediate entity decrypts the information and re-encrypts it before forwarding it · CPC title