Privacy-preserving sanitization for visual computing queries

1. An apparatus, comprising:

a communication interface to communicate with a visual computing device over a network; and

a processor to:

access visual data captured by a camera;

perform a privacy analysis on the visual data using a convolutional neural network (CNN), wherein the CNN is trained to detect a plurality of features, and wherein the CNN comprises:

a plurality of convolutional layers for detecting the plurality of features based on a plurality of convolution filters that activate when convolved with the visual data; and

a stop list of filters indicating a subset of the plurality of convolution filters that correspond to one or more privacy-sensitive features from the plurality of features, wherein the one or more privacy-sensitive features are to be sanitized from the visual data;

detect the one or more privacy-sensitive features in the visual data based on the privacy analysis using the CNN;

sanitize the visual data to mask the one or more privacy-sensitive features, wherein sanitizing the visual data causes sanitized visual data to be produced; and

transmit, via the communication interface, the sanitized visual data to the visual computing device over the network, wherein the visual computing device is to use the sanitized visual data to process a visual query associated with the visual data.

2. The apparatus of claim 1 , wherein the one or more privacy-sensitive features comprise a face of a person.

3. The apparatus of claim 2 , wherein the processor to sanitize the visual data to mask the one or more privacy-sensitive features is further to:

select an avatar to represent the face; and

replace the face with the avatar in the sanitized visual data.

4. The apparatus of claim 3 , wherein the processor to select the avatar to represent the face is further to:

determine that the avatar resembles one or more characteristics of the face.

5. The apparatus of claim 1 , wherein the processor to sanitize the visual data to mask the one or more privacy-sensitive features is further to:

insert random noise into the sanitized visual data to mask the one or more privacy-sensitive features.

6. The apparatus of claim 1 , wherein the processor to sanitize the visual data to mask the one or more privacy-sensitive features is further to:

blur the one or more privacy-sensitive features in the sanitized visual data.

7. The apparatus of claim 1 , wherein the processor to sanitize the visual data to mask the one or more privacy-sensitive features is further to:

identify a set of pixels corresponding to the one or more privacy-sensitive features in the visual data; and

replace the set of pixels with corresponding feature labels in the sanitized visual data, wherein the corresponding feature labels indicate one or more feature types of the one or more privacy-sensitive features corresponding to the set of pixels.

8. A system, comprising:

one or more vision sensors to capture visual data associated with an environment;

a communication interface to communicate with a visual computing device over a network; and

one or more processing devices to:

access the visual data captured by the one or more vision sensors;

perform a privacy analysis on the visual data using a convolutional neural network (CNN), wherein the CNN is trained to detect a plurality of features, and wherein the CNN comprises:

a plurality of convolutional layers for detecting the plurality of features based on a plurality of convolution filters that activate when convolved with the visual data; and

a stop list of filters indicating a subset of the plurality of convolution filters that correspond to one or more privacy-sensitive features from the plurality of features, wherein the one or more privacy-sensitive features are to be sanitized from the visual data;

detect the one or more privacy-sensitive features in the visual data based on the privacy analysis using the CNN;

sanitize the visual data to mask the one or more privacy-sensitive features, wherein sanitizing the visual data causes sanitized visual data to be produced; and

transmit, via the communication interface, the sanitized visual data to the visual computing device over the network, wherein the visual computing device is to use the sanitized visual data to process a visual query associated with the visual data.

9. The system of claim 8 , wherein the one or more vision sensors comprise a camera.

10. The system of claim 8 , wherein the one or more privacy-sensitive features comprise a face of a person.

11. The system of claim 10 , wherein the one or more processing devices to sanitize the visual data to mask the one or more privacy-sensitive features are further to:

select an avatar to represent the face; and

replace the face with the avatar in the sanitized visual data.

12. The system of claim 11 , wherein the one or more processing devices to select the avatar to represent the face are further to:

determine that the avatar resembles one or more characteristics of the face.

13. The system of claim 8 , wherein the one or more processing devices to sanitize the visual data to mask the one or more privacy-sensitive features are further to:

insert random noise into the sanitized visual data to mask the one or more privacy-sensitive features.

14. The system of claim 8 , wherein the one or more processing devices to sanitize the visual data to mask the one or more privacy-sensitive features are further to:

blur the one or more privacy-sensitive features in the sanitized visual data.

15. The system of claim 8 , wherein the one or more processing devices to sanitize the visual data to mask the one or more privacy-sensitive features are further to:

identify a set of pixels corresponding to the one or more privacy-sensitive features in the visual data; and

replace the set of pixels with corresponding feature labels in the sanitized visual data, wherein the corresponding feature labels indicate one or more feature types of the one or more privacy-sensitive features corresponding to the set of pixels.

16. At least one non-transitory machine accessible storage medium having instructions stored thereon, wherein the instructions, when executed on a machine, cause the machine to:

access visual data captured by a camera;

perform a privacy analysis on the visual data using a convolutional neural network (CNN), wherein the CNN is trained to detect a plurality of features, and wherein the CNN comprises:

a plurality of convolutional layers for detecting the plurality of features based on a plurality of convolution filters that activate when convolved with the visual data; and

a stop list of filters indicating a subset of the plurality of convolution filters that correspond to one or more privacy-sensitive features from the plurality of features, wherein the one or more privacy-sensitive features are to be sanitized from the visual data;

detect the one or more privacy-sensitive features in the visual data based on the privacy analysis using the CNN;

sanitize the visual data to mask the one or more privacy-sensitive features, wherein sanitizing the visual data causes sanitized visual data to be produced; and

transmit the sanitized visual data to a visual computing device over a network, wherein the visual computing device is to use the sanitized visual data to process a visual query associated with the visual data.

17. The storage medium of claim 16 , wherein:

the one or more privacy-sensitive features comprise a face of a person; and

the instructions that cause the machine to sanitize the visual data to mask the one or more privacy-sensitive features further cause the machine to:

select an avatar to represent the face; and

replace the face with the avatar in the sanitized visual data.

18. The storage medium of claim 16 , wherein the instructions that cause the machine to sanitize the visual data to mask the one or more privacy-sensitive features further cause the machine to:

insert random noise into the sanitized visual data to mask the one or more privacy-sensitive features.

19. The storage medium of claim 16 , wherein the instructions that cause the machine to sanitize the visual data to mask the one or more privacy-sensitive features-further cause the machine to:

identify a set of pixels corresponding to the one or more privacy-sensitive features in the visual data; and

replace the set of pixels with corresponding feature labels in the sanitized visual data, wherein the corresponding feature labels indicate one or more feature types of the one or more privacy-sensitive features corresponding to the set of pixels.

20. A method, comprising:

accessing visual data captured by a camera;

performing a privacy analysis on the visual data using a convolutional neural network (CNN), wherein the CNN is trained to detect a plurality of features, and wherein the CNN comprises:

a plurality of convolutional layers for detecting the plurality of features based on a plurality of convolution filters that activate when convolved with the visual data; and

a stop list of filters indicating a subset of the plurality of convolution filters that correspond to one or more privacy-sensitive features from the plurality of features, wherein the one or more privacy-sensitive features are to be sanitized from the visual data;

detecting the one or more privacy-sensitive features in the visual data based on the privacy analysis using the CNN;

sanitizing the visual data to mask the one or more privacy-sensitive features, wherein sanitizing the visual data causes sanitized visual data to be produced; and

transmitting the sanitized visual data to a visual computing device over a network, wherein the visual computing device is to use the sanitized visual data to process a visual query associated with the visual data.

21. The method of claim 20 , wherein:

the one or more privacy-sensitive features comprise a face of a person; and

sanitizing the visual data to mask the one or more privacy-sensitive features comprises:

selecting an avatar to represent the face; and

replacing the face with the avatar in the sanitized visual data.

22. The method of claim 20 , wherein sanitizing the visual data to mask the one or more privacy-sensitive features comprises:

inserting random noise into the sanitized visual data to mask the one or more privacy-sensitive features.

23. The method of claim 20 , wherein sanitizing the visual data to mask the one or more privacy-sensitive features comprises:

identifying a set of pixels corresponding to the one or more privacy-sensitive features in the visual data; and

replacing the set of pixels with corresponding feature labels in the sanitized visual data, wherein the corresponding feature labels indicate one or more feature types of the one or more privacy-sensitive features corresponding to the set of pixels.