Neural network digital signature encoding method, device and corresponding computer program
AssigneeThales
InventorsHenri Belfy, Tom Fougere
Life of this patent
Estimating the term…
Abstract
A method for encoding a digital signature of a neural network, implemented by an electronic device, the neural network being stored within a data structure including blocks of parameters. The method includes, for a current block of parameters including at least M parameters representing real numbers: an operation of acquiring a signature binary image from the digital signature including a set of bits extracted from parameters of the current block of parameters, and an operation of combining the signature binary image with a reference binary image delivering a coded binary image.
Claims12
An encoding method, implemented by an electronic device, encoding a digital signature of a neural network, the neural network being stored within a data structure, the neural network comprising blocks of neural network parameters, wherein a current block of neural network parameters comprises at least M neural network parameters representing real numbers, the neural network parameters having been modified during a training of the neural network and being selected from the group consisting of:
- layer weights, biases, tensor values, normalization values, and convolution values, the encoding method comprising:
- acquiring a signature binary image from the digital signature of the neural network, the digital signature comprising a set of bits extracted from the at least M neural network parameters, and the signature binary image being an array of binary pixels;
- combining the signature binary image with a reference binary image for delivering an encoded binary image;
- and
- using the encoded binary image to authenticate the neural network even when the neural network is modified by an attacker.
An electronic device comprising an encoder encoding a digital signature of a neural network, the neural network being stored within a data structure, the neural network comprising blocks of neural network parameters, wherein a current block of neural network parameters comprises at least M neural network parameters representing real numbers, the neural network parameters having been modified during a training of the neural network and being chosen from the group consisting of:
- layer weights, biases, tensor values, normalization values and convolution values, the encoder comprising:
- means for acquiring a signature binary image from the digital signature of the neural network, the digital signature comprising a set of bits extracted from the at least M neural network parameters, and the signature binary image being an array of binary pixels;
- means for combining the signature binary image with a reference binary image for delivering an encoded binary image;
- and
- means for using the encoded binary image to authenticate the neural network even when the neural network is modified by an attacker.
A decoding method, implemented by an electronic device, decoding a digital signature of a neural network, the neural network being stored within a data structure, the neural network comprising blocks of neural network parameters, wherein a current block of neural network parameters comprises at least M neural network parameters representing real numbers, the neural network parameters having been modified during a training of the neural network and being selected from the group consisting of:
- layer weights, biases, tensor values, normalization values, and convolution values, the method comprising:
- acquiring a signature binary image from the signature of the neural network, the signature comprising a set of bits extracted from the at least M neural network parameters, and the signature binary image being an array of binary pixels;
- combining the signature binary image with a coded binary image associated with the current block of parameters, for delivering an induced reference binary;
- and
- using the induced reference binary image to authenticate the neural network even when the neural network is modified by an attacker.
An electronic device comprising a decoder decoding a digital signature of a neural network, the neural network being stored within a data structure, the neural network comprising blocks of neural network parameters, wherein a current block of neural network parameters comprises at least M neural network parameters representing real numbers, the neural network parameters having been modified during a training of the neural network and being chosen from the group consisting of:
- layer weights, biases, tensor values, normalization values and convolution values, the decoder comprising:
- means for acquiring a signature binary image from a set of parameter bits of the at least M neural network parameters, the signature binary image being an array of binary pixels;
- means for combining the signature binary image with a coded binary image associated with the current block of parameters, for delivering an induced reference binary image;
- and
- means for using the induced reference binary image to authenticate the neural network even when the neural network is modified by an attacker.
Dependencies are inferred from the claim text.