Machine learning-based early detection of malicious G-code manipulations in 3D printing

The increased adoption of 3D printing across various critical manufacturing sectors has made it a fruitful target for adversaries, particularly through the manipulation of G-code instructions that control the operations of 3D printers. Simple modifications to these instructions could significantly impact the integrity of 3D-printed objects. While side-channel analysis during printing is a common detection method, identifying potential malicious G-code before printing can save time and resources. Existing work relies on primitive encryption and hashing techniques and cannot distinguish between benign and malicious G-code instructions. It assumes that G-code files are benign and uses them as a reference model, focusing only on the integrity checking of G-code during storage and transmission. This paper introduces a novel automated approach to efficiently differentiate between benign and subtly manipulated G-code caused by filament, thermodynamic, and Z-profile attacks without requiring a reference model. As the first study leveraging recent advancements in Machine Learning (ML), we address several challenges in dataset generation, feature engineering, G-code segmenting and labeling, and ML classifier selection. We generate diverse G-code datasets to identify the optimal dataset characteristics and conduct a comprehensive formal analysis to extract the most suitable features. Efficient labeling strategies are employed at both layer and command levels, using the Multiple Instance Learning (MIL) paradigm for the former. We adopt the Bidirectional Long Short-Term Memory (Bi-LSTM) model enhanced by an attention mechanism and focal loss function for layer classification. Meanwhile, the Random Forest (RF) algorithm and Multilayer Perceptron (MLP) neural network model are used for command classification. All classifiers are designed to handle the imbalanced dataset. Experimental evaluation demonstrates the efficacy of our approach. The Bi-LSTM model achieves F1 scores up to 91.3% in detecting filament attacks, while the RF algorithm performs better in detecting nuanced thermodynamic and Z-profile changes at the command level, achieving F1 scores between 81.6% and 99.3%.