Methods for secure learning of parameters of a convolution neural network, and for secure input data classification
A method for secure learning of parameters of a convolution neural network, CNN, for data classification includes the implementation, by data processing of a first server, including receiving from a second server a base of already classified learning data, the learning data being homomorphically encrypted; learning in the encrypted domain, from the learning database, the parameters of a reference CNN including a non-linear layer (POLYNOMIAL) operating an at least two-degree polynomial function approximating an activation function; a batch normalization layer before each non-linear layer (POLYNOMIAL); and transmitting the learnt parameters to the second server, for decryption and use for classification.
1. A method for secure learning of parameters of a convolution neural network, CNN, for input data classification, the method comprising the implementation, by data processing means of a first server, of steps of:
receiving from a second server a base of already classified learning data, said learning data being homomorphically encrypted;
learning in an encrypted domain, from said base of already classified learning data, the parameters of a reference CNN comprising at least:
a non-linear layer (POLYNOMIAL) operating an at least two-degree polynomial function approximating an activation function, and
a non-approximated batch normalization layer before each non-linear layer (POLYNOMIAL); and
transmitting the learnt parameters to said second server, for decryption and use for classification.
2. The method according to claim 1 , wherein said polynomial function is determined before learning via polynomial regression of said activation function from points randomly selected in a given distribution.
3. The method according to claim 1 , wherein said polynomial function is determined during learning, the coefficients of said at least two-degree polynomial function belonging to the learnt parameters.
4. The method according to claim 1 , wherein the reference CNN comprises a convolution layer before each batch normalization layer.
5. The method according to claim 1 , wherein the reference CNN comprises at least one pooling layer operating a function of (AVERAGEPOOL) type, after a non-linear layer (POLYNOMIAL).
6. The method according to claim 1 , wherein the reference CNN comprises at least one final, fully connected layer.
7. The method according to claim 4 , wherein the reference CNN comprises at least one pooling layer operating a function of (AVERAGEPOOL) type, after a non-linear layer (POLYNOMIAL), wherein the reference CNN comprises at least one final, fully connected layer, and wherein the architecture of the reference CNN is p →AVERAGEPOOL] n →FC→FC or p →AVERAGEPOOL] n →FC →FC.
8. The method according to claim 1 , wherein said activation function is of Rectified Linear Unit type, ReLU.
9. The method according to claim 1 , wherein said polynomial function is of degree two or three, preferably two.
10. The method according to claim 1 , wherein said input data or learning data represent images, said classification being object recognition.
11. A non-transitory storage means readable by computer equipment on which a computer programme product is installed comprising code instructions for execution of the method according to claim 1 for the secure learning of parameters of a convolution neural network CNN.
12. A method for secure input data classification, comprising:
learning, by data processing means of a first server from a base of already classified learning data, parameters of a reference convolution neural network CNN comprising at least:
a non-linear layer (POLYNOMIAL) operating an at least two-degree polynomial function approximating an activation function, and
a non-approximated batch normalization layer before each non-linear layer (POLYNOMIAL);
receiving said input data, homomorphically encrypted, by data processing means of a second server, from client equipment;
classifying said encrypted input data in an encrypted domain by the data processing means of the second server, with the reference CNN, so as to obtain an encrypted classification result; and
transmitting the obtained encrypted classification result to said client equipment, for decryption.
13. The method according to claim 12 , wherein said learning includes receiving from the second server the base of already classified learning data, said learning data being homomorphically encrypted, and the learning is performed in the encrypted domain.
14. A method for secure input data classification, comprising:
learning, by data processing means of a first server from a base of already classified learning data, parameters of a reference convolution neural network, CNN, comprising at least:
a non-linear layer operating an activation function, and
a non-approximated batch normalization layer before each non-linear layer,
the learning comprising the determination of an at least two-degree polynomial function approximating said activation function;
receiving said input data, homomorphically encrypted, by data processing means of a second server, from client equipment;
classifying said encrypted input data in an encrypted domain, by the data processing means of the second server, with a substitution CNN using the parameters learnt for the reference CNN and comprising, instead of each non-linear layer operating the activation function, a non-linear layer (POLYNOMIAL) operating said determined at least two-degree polynomial function, so as to obtain an encrypted classification result; and
transmitting said obtained encrypted classification result to said client equipment, for decryption.
15. The method according to claim 14 , wherein said polynomial function is determined via polynomial regression of said activation function from points randomly selected in a given distribution.
16. The method according to claim 14 , wherein said polynomial function is determined via polynomial regression of said activation function from points recovered at input of one or more non-linear layers of the reference CNN.
17. The method according to claim 14 , wherein the learning, after determination of the parameters of the reference CNN, comprises the implementation of at least one additional iteration of learning on the substitution CNN to adapt said parameters to said determined polynomial function.