Defending neural networks by randomizing model weights

A neural network defense method, comprising:

• generating, by random number generation circuitry, a plurality of pseudo-random values between a defined first boundary value and a defined second boundary value responsive to receipt of an initial seed value;
• and
• combining, by combiner circuitry coupled to the random number generation circuitry, each of a plurality of weights associated with a trained neural network model with a respective one of the plurality of pseudo-random values to provide a plurality of stochastic neural network weight values;
• wherein each of the plurality of stochastic neural network weight values have a defined number of randomized least significant mantissa bits/digits.

A non-transitory storage device including instructions that, when executed by processor circuitry, cause the processor circuitry to:

• generate a plurality of pseudo-random values between a defined first boundary value and a defined second boundary value responsive to receipt of an initial seed value;
• and
• combine each of a plurality of weights associated with a trained neural network model with a respective one of the plurality of pseudo-random values to provide a plurality of stochastic neural network weight values;
• wherein each of the plurality of stochastic neural network weight values have a defined number of randomized least significant mantissa bits/digits.

A neural network defense system, comprising:

• means for generating a plurality of pseudo-random values between a defined first boundary value and a defined second boundary value responsive to receipt of an initial seed value;
• and
• means for combining each of a plurality of weights associated with a trained neural network model with a respective one of the plurality of pseudo-random values to provide a plurality of stochastic neural network weight values;
• wherein each of the plurality of stochastic neural network weight values have a defined number of randomized least significant mantissa bits/digits.