Precision Machine Learning

Machine learning cryptography methods for IoT in healthcare

BACKGROUND

The increased application of Internet of Things (IoT) in healthcare, has fueled concerns regarding the security and privacy of patient data. Lightweight Cryptography (LWC) algorithms can be seen as a potential solution to address this concern. Due to the high variation of LWC, the primary objective of this study was to identify a suitable yet effective algorithm for securing sensitive patient information on IoT devices.

METHODS

This study evaluates the performance of eight LWC algorithms-AES, PRESENT, MSEA, LEA, XTEA, SIMON, PRINCE, and RECTANGLE-using machine learning models. Experiments were conducted on a Raspberry Pi 3 microcontroller using 16 KB to 2048 KB files. Machine learning models were trained and tested for each LWC algorithm and their performance was evaluated based using precision, recall, F1-score, and accuracy metrics.

RESULTS

The study analyzed the encryption/decryption execution time, energy consumption, memory usage, and throughput of eight LWC algorithms. The RECTANGLE algorithm was identified as the most suitable and efficient LWC algorithm for IoT in healthcare due to its speed, efficiency, simplicity, and flexibility.

CONCLUSIONS

This research addresses security and privacy concerns in IoT healthcare and identifies key performance factors of LWC algorithms utilizing the SLR research methodology. Furthermore, the study provides insights into the optimal choice of LWC algorithm for enhancing privacy and security in IoT healthcare environments.

Approximating the nuclear binding energy using analytic continued fractions

Understanding nuclear behaviour is fundamental in nuclear physics. This paper introduces a data-driven approach, Continued Fraction Regression (cf-r), to analyze nuclear binding energy (B(A, Z)). Using a tailored loss function and analytic continued fractions, our method accurately approximates stable and experimentally confirmed unstable nuclides. We identify the best model for nuclides with A ≥ 200 , achieving precise predictions with residuals smaller than 0.15 MeV. Our model's extrapolation capabilities are demonstrated as it converges with upper and lower bounds at the nuclear mass limit, reinforcing its accuracy and robustness. The results offer valuable insights into the current limitations of state-of-the-art data-driven approaches in approximating the nuclear binding energy. This work provides an illustration on the use of analytical continued fraction regression for a wide range of other possible applications.

Continued fractions and the Thomson problem

We introduce new analytical approximations of the minimum electrostatic energy configuration of n electrons, E(n), when they are constrained to be on the surface of a unit sphere. Using 453 putative optimal configurations, we searched for approximations of the form [Formula: see text] where g(n) was obtained via a memetic algorithm that searched for truncated analytic continued fractions finally obtaining one with Mean Squared Error equal to [Formula: see text] for the model of the normalized energy ([Formula: see text]). Using the Online Encyclopedia of Integer Sequences, we searched over 350,000 sequences and, for small values of n, we identified a strong correlation of the highest residual of our best approximations with the sequence of integers n defined by the condition that [Formula: see text] is a prime. We also observed an interesting correlation with the behavior of the smallest angle [Formula: see text], measured in radians, subtended by the vectors associated with the nearest pair of electrons in the optimal configuration. When using both [Formula: see text] and [Formula: see text] as variables a very simple approximation formula for [Formula: see text] was obtained with MSE= [Formula: see text] and MSE= 73.2349 for E(n). When expanded as a power series in infinity, we observe that an unknown constant of an expansion as a function of [Formula: see text] of E(n) first proposed by Glasser and Every in 1992 as [Formula: see text], and later refined by Morris, Deaven and Ho as [Formula: see text] in 1996, may actually be very close to -1.10462553440167 when the assumed optima for [Formula: see text] are used.