Artificial neural network integrity verification

A method, comprising:

• storing, in a memory of a system employing an artificial neural network (ANN) comprising a plurality of ANN partitions, cryptographic codes corresponding to the plurality of ANN partitions;
• wherein a particular ANN partition of the plurality of ANN partitions has a corresponding set of synaptic weights and biases stored in the memory;
• wherein a cryptographic code corresponding to the particular ANN partition comprises a digest that is a hash of the corresponding set of synaptic weights and biases stored in the memory;
• performing an integrity verification operation on the particular ANN partition by:
• reading, from the memory:
• the corresponding set of synaptic weights and biases;
• and
• the cryptographic code corresponding to the particular ANN partition;
• generating, using the set of synaptic weights and biases read from the memory, a verification cryptographic code corresponding to the particular ANN partition, wherein the verification cryptographic code comprises a verification digest that is a hash of the set of synaptic weights and biases read from the memory;
• comparing the digest corresponding to the particular ANN partition with the verification digest;
• determining that the digest corresponding to the particular ANN partition is different than the verification digest;
• and
• in response to determining that the digest corresponding to the particular ANN is different than the verification digest, providing an indication of an integrity verification failure;
• receiving updates to the corresponding set of synaptic weights and biases corresponding to the particular ANN partition;
• and in response to receiving the updates to the corresponding set of synaptic weights and biases, performing an additional integrity verification operation, wherein the additional integrity verification operation results in an integrity verification pass or an integrity verification fail;
• in response to the additional integrity verification operation resulting in the integrity verification pass:
• replacing the corresponding set of synaptic weights and biases stored in the memory and corresponding to the particular ANN with the updates to the corresponding set of synaptic weights and biases;
• generating an updated cryptographic code corresponding to the particular ANN partition based on the updates to the corresponding set of synaptic weights and biases;
• and
• storing the updated cryptographic code in the memory;
• and
• in response to the additional integrity verification operation resulting in the integrity verification fail, preventing the replacing of the corresponding set of synaptic weights and biases stored in the memory and corresponding to the particular ANN with the updates to the corresponding set of synaptic weights and biases.

An apparatus employing an artificial neural network (ANN), the apparatus comprising:

• a memory resource configured to:
• store sets of synaptic weights and biases corresponding to ANN partitions;
• store cryptographic codes corresponding to the ANN partitions, wherein the cryptographic codes comprise digests generated by a hash on the sets of synaptic weights and biases corresponding to the ANN partitions;
• and
• a processor configured to, in association with performing an integrity verification operation on a particular ANN partition:
• read, from the memory resource:
• a set of synaptic weights and biases corresponding to the particular ANN partition;
• and
• a cryptographic code corresponding to the particular ANN partition;
• generate, using the set of synaptic weights and biases read from the memory resource, a verification cryptographic code corresponding to the particular ANN partition, wherein the verification cryptographic code comprises a verification digest generated by performing a hash on the set of synaptic weights and biases read from the memory resource and corresponding to the particular ANN partition;
• compare the digest corresponding to the particular ANN partition with the verification digest;
• determine that the digest corresponding to the particular ANN partition is different than the verification digest;
• and
• in response to determining that the digest corresponding to the particular ANN partition is different than the verification digest, provide an indication of an integrity verification failure;
• wherein the processor is further configured to:
• receive updates to the set of synaptic weights and biases corresponding to the particular ANN partition;
• and in response to receiving the updates to the set of synaptic weights and biases, perform an additional integrity verification operation, wherein the additional integrity verification operation results in an integrity verification pass or an integrity verification fail;
• in response to the additional integrity verification operation resulting in the integrity verification pass:
• replace the set of synaptic weights and biases stored in the memory and corresponding to the particular ANN with the updates to the set of synaptic weights and biases;
• generate an updated cryptographic code corresponding to the particular ANN partition based on the updates to the set of synaptic weights and biases;
• and
• store the updated cryptographic code in the memory;
• and
• in response to the additional integrity verification operation resulting in the integrity verification fail, prevent the replacing of the set of synaptic weights and biases stored in the memory and corresponding to the particular ANN with the updates to the set of synaptic weights and biases.

An apparatus employing an artificial neural network (ANN) having a plurality of ANN partitions, the apparatus comprising:

• a memory resource configured to store weight matrix data and bias data corresponding to the plurality of ANN partitions;
• and
• a processor configured to:
• receive a number of inputs;
• determine a particular ANN partition of the plurality of ANN partitions to which the number of inputs correspond;
• read, from the memory resource:
• weight matrix data and bias data corresponding to the particular ANN partition;
• and
• a first cryptographic signature corresponding to the particular ANN partition;
• generate, using the weight matrix data and bias data read from the memory, a second cryptographic signature corresponding to the particular ANN partition;
• determine whether the first cryptographic signature and the second cryptographic signature match;
• determine a mismatch between the first cryptographic signature and the second cryptographic signature;
• and
• responsive to determining the mismatch between the first cryptographic signature and the second cryptographic signature, issue an indication of the mismatch;
• and
• wherein the processor is further configured to:
• receive updated weight matrix data and bias data corresponding to the particular ANN partition;
• and in response to receiving the updated weight matrix data and bias data, perform an integrity verification operation, wherein the integrity verification operation results in an integrity verification pass or an integrity verification fail;
• in response to the integrity verification operation resulting in the integrity verification pass:
• replace the weight matrix data and bias data stored in the memory resource and corresponding to the particular ANN with the updated weight matrix data and bias data;
• generate an updated cryptographic signature corresponding to the particular ANN partition based on the updated weight matrix data and bias data;
• and
• store the updated cryptographic signature in the memory resource;
• and
• in response to the integrity verification operation resulting in the integrity verification fail, prevent the replacing of the weight matrix data and bias data stored in the memory resource and corresponding to the particular ANN with the updated weight matrix data and bias data.