Machine Learning Assisted Accurate Estimation of Usage Duration and Manufacturer for Recycled and Counterfeit Flash Memory Detection

Abstract

With the large-scale adaptation of a horizontal business model, modern semiconductor supply chain is plagued by recycled and counterfeit ICs, including flash memory chips. Since flash memory modules have an inherently finite lifespan, detection of recycled flash memory chips before their deployment in safety-critical systems is important to prevent disastrous consequences. The state-of-art detection methods can detect flash memory modules between 0.05% to 3.00% of their lifespan as minimum usage duration, depending on the details of the flash memory chip. In this paper, we propose a versatile machine learning assisted detection methodology to improve the minimum usage duration accuracy between 0.05% to 0.96% of their lifespan, and also to accurately associate a flash memory IC with its manufacturer. Through detailed experimentation and comparison of detection results obtained using three popular supervised machine learning techniques (Support Vector Machines, Logistic Regression and Artificial Neural Networks), we demonstrate that usage of features composed of multiple characteristics of a given chip, rather than just a single property of a chip (as used in previous works), improves detection accuracy.