Optimizations for Learning from Linear Feedback Shift Register Variations with Artificial Neural Networks

Recent years have seen an increase in Machine Learning techniques being applied to learn from Pseudorandom Number Generators (PRNGs). Currently, the best results have been obtained for learning on Linear Feedback Shift Registers (LFSRs). Due to the deterministic nature of LFSRs, Decision Trees (DTs) and Artificial Neural Networks (ANNs) were able to reach up to \(100\%\) test accuracy for next bit prediction tasks. Despite important advances, a number of directions have been neglected. The current work sets to investigate such directions and bring a more comprehensive understanding of the ANNs capabilities in this context, cementing them as the most reliable technique for learning on LFSRs and their variations. More precisely, the study presents the results of learning from the previously uninvestigated Galois form a LFSR. Moreover, an optimization is proposed with respect to the number of bits needed for learning to predict the outputs of the Geffe generator through the introduction of an ANN pipeline model. Performed experiments display the strength of the approach that is able to maintain up to \(100\%\) accuracy for predicting Geffe outputs while reducing the amount of training bits to the degree of magnitude \(10^3\) for each LFSR in the proposed pipeline.