Brain Programming and Its Resilience Using a Real-World Database of a Snowy Plover Shorebird Conference Paper uri icon

abstract

  • Even when deep convolutional neural networks have proven to be effective at saliency detection, they have a vulnerability that should not be ignored: they are susceptible to adversarial attacks, making them highly unreliable. Reliability is an important aspect to consider when it comes to salient object detection; without it, an attacker can render the algorithm useless. Brain programming–an evolutionary methodology for visual problems–is highly resilient and can withstand even the most intense perturbations. In this work, we perform for the first time a study that compares the resilience against adversarial attacks and noise perturbations using a real-world database of a shorebird called the Snowy Plover in a visual attention task. Database images were taken on the field and even posed a detection challenge due to the nature of the environment and the bird’s physical characteristics. By attacking three different deep convolutional neural networks using adversarial examples from this database, we prove that they are no match for the brain programming algorithm when it comes to resilience, suffering significant losses in their performance. On the other hand, brain programming stands its ground and sees its performance unaffected. Also, by using images of the Snowy Plover, we refer to the importance of resilience in real-world issues where conservation is present. Brain programming is the first highly resilient evolutionary algorithm used for saliency detection tasks. © 2022, Springer Nature Switzerland AG.

publication date

  • 2022-01-01