Laser-based holography for projecting adversarial perturbations to attack image classifiers

This paper presents two complementary methods, the Wavelength-Specific Map Attack (WSMA) and Neural Beam Fooling (NBF), that translate digital adversarial techniques into the physical realm using laser-based holography and a digital micromirror device (DMD). WSMA is an adversarial attack that extends the Jacobian-based Saliency Map Attack (JSMA) by optimizing perturbation placement based on the wavelength of the perturbation laser and generalizing its approach to operate on any gradient-based attribution map. NBF enables the physical realization of these perturbations by leveraging a spatial light modulator and the Gerchberg-Saxton algorithm to project structured light patterns that induce targeted misclassification in real-world camera systems. Additionally, NBF incorporates a novel adaptive filtering step to suppress twin-image artifacts and enhance projection fidelity. We demonstrate the successful deployment of sparse adversarial perturbations under varying conditions, highlighting the feasibility of real-world adversarial interference and offering new insights into bridging digital and physical attack techniques.