Image-Scaling Attack Prevention in Industrial Cyber Physical Systems

The rise of the Internet of Things movement has made machine-learning (ML) pipelines a staple of industrial cyber-physical systems (ICPSs), and a new attack surface. Image-scaling backdoors, inserted during data acquisition or retraining, can silently mislead these pipelines. We examine this threat on five representative product classes from the MVTecAD dataset and introduce two lightweight, pre-processing prevention mechanisms. The first couples a low-threshold Canny edge detector with a decision tree, flagging poisoned frames by their abnormally dense corner maps and reaching 98 % accuracy on unseen images. The second compares colour-histogram peaks after a downscale–upscale circuit, achieving 88 % accuracy. Both filters are deterministic, sensor-side, and incur negligible overhead, with the Canny-based variant offering the highest precision for real-time ICPS deployment.