Robust image classification against adversarial attacks using elastic similarity measures between edge count sequences

Izaskun Oregi, Javier Del Ser, Aritz Pérez, José Antonio Lozano

Abstract: Highlights•We propose a novel defense strategy for image classifiers against adversarial attacks.•The defense strategy extracts color based sequences from the images.•The sequence legitimacy is analyzed by kernel models used in time series analysis.•Results confirm the increased robustness of the classifier provided by our strategy.