ML Inference Obfuscation for Side-Channel Parameter Protection

The patent introduces a random number generator as an intermediary component that generates random values to mask the actual neural network operations. This random number generator acts as a mediator between the input data and the neural network processing, inserting randomness into the computational path without requiring fundamental changes to the hardware architecture. The random values are combined with actual data during processing, making it difficult for attackers to correlate power consumption or electromagnetic emissions with specific neural network operations, thereby preventing side channel attacks while maintaining hardware feasibility.

The patent merges the obfuscating operations with the inference operations by integrating the random number generation and masking processes into the existing neural network computation pipeline. Instead of performing separate obfuscating operations that would double the computational workload, the random values are generated and combined concurrently with the actual inference computations. This merging approach allows the security-enhancing operations to share computational resources and processing time with the inference operations, reducing the overall computational overhead while maintaining security.

The patent changes the parameters of the neural network operations by introducing random values that alter the power consumption and electromagnetic emission characteristics during computation. By varying the data values with random noise, the physical side channels (power consumption, electromagnetic emissions) that attackers monitor are no longer directly correlated with the actual neural network operations. This parameter change effectively masks the true computational patterns, making side channel measurements inaccurate for attack purposes while preserving the functional correctness of the neural network output.