Vulnerabilities and robustness in computer vision

Summary

The dissertation intends to examine the reliability and robustness of the most recent computer vision models in environments different from those they have been trained on. The study will focus on the performance of the models on idiosyncratic datasets and in environments with malicious users. Specifically, the research phases will include the creation of multiple state-of-the-art computer vision models with different architectures, and after the verification of their performance on common datasets, we will proceed to test them on idiosyncratic datasets, such as ObjectNet, while also examining their resilience to black and white box adversarial attacks. Based on the results of these tests, we will evaluate the effectiveness, reliability, and robustness of these computer vision models.

Additionally, we will examine the transferability of some of these attacks among different model architectures. This approach will allow the identification of potential weaknesses in the models' ability to generalize their knowledge to uncontrolled and adversarial environments and open the discussion for possible defenses and mitigation measures for these weaknesses, as well as the capabilities of each architecture. Variants in neural networks and attacks will be selected based on the specific needs of the research and the continuous updates in the field.