In-Memory Zero-Space Floating-Point-Based CNN Protection Using Non-significant and Invariant Bits

Convolutional Neural Networks (CNNs) have accomplished significant success in various domains, including transportation, health care and banking. Millions of weights, loaded from main memory into the internal buffers of CNN accelerators, are repeatedly used in the inference process. Accidental and malicious bit-flips targeting these buffers may negatively impact the CNN’s accuracy. This paper proposes a methodology to tolerate the effect of (multiple) bit-flips on floating-point-based CNNs using the non-significant and the invariant bits of CNN parameters. The former, determined after fault injection, do not significantly affect the accuracy of the inference process regardless of their value. The latter, determined after analyzing the network parameters, have the same value for all of them. Slight modifications can be applied to carefully selected parameters to increase the number of invariant bits. Since non-significant and invariant bits do not require protection against faults, they are employed to store the parity bits of error control codes. The methodology preserves the CNN accuracy, keeps its memory footprint, and does not require any retraining. Its usefulness is exemplified through the FP32 and BFloat16 versions of the LeNet-5 and GoogleNet CNNs.