Method for collaborative learning of an artificial neural network without disclosing training data

A method for federative learning of a model of an artificial neural network, the model being characterised by parameters giving synaptic coefficients of said network, the method making use of a plurality of N data suppliers (DS1, . . . , DSN) each having a distinct set of learning data (S1, . . . , SN) and an aggregation platform (FP), said method beginning with an initialisation step to initialise the parameters of the model, and proceeding with a plurality of successive iterations, each iteration updating the parameters of a current model of the neural network, wherein:

• in the initialisation step, the data suppliers share the private key and the public key of a homomorphic cryptosystem (HE.sk,HE.pk), each data supplier also generating a secret stream encryption key, Ksymn, and transmitting the public key of the homomorphic cryptosystem encrypted using said secret key, Enc(Ksymn;HE.pk) to the aggregation platform;
• and wherein each iteration t,
• the aggregation platform selects a sub-plurality M≤N of data suppliers and transmits the parameters of the current model, homomorphically encrypted, Enc(wt,HE.pk), to the M data suppliers;
• each selected data supplier decrypts the parameters thus encrypted using the private key of the homomorphic cryptosystem, trains the current model on its learning data set to obtain a partial model;
• each selected data supplier, DSn, encrypts) the parameters of the partial model using his secret key, Ksymn, and transmits the parameters of the partial model thus encrypted, Enc(wnt,Ksymn), to the aggregation platform;
• the aggregation platform transcrypts parameters of the partial models to the different selected data suppliers, to obtain these same parameters encrypted in the homomorphic domain,
• the aggregation platform combines, in the homomorphic domain, parameters of the partial models obtained by the different data suppliers, to obtain the parameters of a new current model encrypted in the homomorphic domain;
• and, when a stop criterion is satisfied, each data supplier decrypts the parameters of the last current model using the private key of the homomorphic cryptosystem, to obtain a final model, trained on the reunion of said learning data sets.