Description
The increasing adoption of Additive Manufacturing (AM) in critical applications has raised
concerns about process security and integrity assurance. This thesis investigates multiple
single-modal Side Channel (SC) monitoring approaches for detecting anomalies in Fused
Filament Fabrication (FFF) 3D printing systems. We develop and evaluate kinetic, acoustic,
optical, and electrical SC methods to reconstruct printer trajectories and identify malicious or
faulty process deviations. For kinetic monitoring, we examine heterogeneous sensor fusion
approaches using Inertial Measurement Unit (IMU), magnetometer, and Time-of-Flight (ToF)
sensors, implementing both direct trajectory reconstruction using ToF distance measurements
and global magnetic field mapping approaches with Gaussian Process regression for printbed
localization. These investigations reveal fundamental trade-offs between measurement invasiveness
and reconstruction accuracy. Optical SC analysis employing marker-based tracking
with an adaptive Kalman Filter (KF) achieves robust trajectory reconstruction, providing
sufficient data quality to enable accurate anomaly detection, using Temporal Convolutional
Network Autoencoders (TCNAEs) for anomaly detection. We achieved Area Under the
Receiver Operating Characteristic (AUROC) values of 0.95 despite the challenging shapes
used for training and testing, with detection sensitivity extending to speed anomalies as brief
as one second within 25 minute print sequences. All experiments are conducted under regular
operating conditions without assumptions that artificially enhance trajectory reconstruction
and detection performance, with environmental disturbances and real-world noise characteristics
preserved in the experimental protocol. For acoustic SC monitoring, we develop an
audio fingerprinting framework using Kalman-filtered Mel-Frequency Cepstral Coefficients
(MFCCs) as similarity metrics and Dynamic Time Warping (DTW) alignment on normalized
vertical amplitude power to detect process deviations relative to master reference profiles. Our
findings demonstrate that SC monitoring provides complementary security layer to traditional
file-based integrity checks, with detection sensitivity sufficient to identify strategically-placed
anomalies that could compromise mechanical properties through altered material deposition
patterns. Critically, our anomaly detection approaches operate independently of printer
instruction files, learning nominal process behavior solely from SC observations rather than
relying on G-code ground truth, enabling detection even when attack vectors compromise
both the instruction stream and physical execution. This work thus lays the foundation for a
robust additional security layer in AM environments, providing complementary monitoring
capabilities that can detect both accidental process deviations and deliberate attack vectors
such as G-code manipulation or firmware compromise.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching