Effect of spatial speckle mixing on data poisoning vulnerability in deep-learning-based scattering imaging

Deep-learning-based imaging through scattering media has attracted considerable attention as a method for reconstructing objects from complex speckle patterns. However, the vulnerability of such systems to training-stage data poisoning attacks has not been sufficiently investigated. In this study, we examine how the optical configuration affects poisoning vulnerability in deep-learning-based scattering imaging. Two representative optical configurations, a conventional imaging configuration and a lensless configuration, were compared experimentally. Trigger patterns were embedded in a subset of the training images and paired with attacker-defined target images to induce backdoor behavior. The results showed that the imaging configuration was vulnerable to poisoning attacks, whereas the lensless configuration exhibited substantially lower attack success rates under the same experimental conditions. This difference is likely related to the spatial distribution of speckle information. In the imaging configuration, trigger-related features remained relatively localized, making them easier for the neural network to associate with incorrect reconstructions. In contrast, in the lensless configuration, object and trigger information were more broadly mixed over the sensor plane, which made trigger learning more difficult. These results show that the spatial distribution of optical information affects poisoning vulnerability in scattering imaging systems.