Transferability of Adversarial Attacks on Tiny Deep Learning Models for IoT Unmanned Aerial Vehicles
Shan Zhou, Xianting Huang, Mohammad S. Obaidat, Bander Alzahrani, Xuming Han, Saru Kumari, Chien‐Ming Chen
Abstract
In the realm of miniature machine learning for Internet of Unmanned Aerial Vehicles (UAVs), the security concerns of machine learning models are obvious, especially when it comes to adversarial attacks. Models can become confused and their performance undermined by the introduction of meticulously crafted distortions. These attacks can even infiltrate a variety of models, bringing greater security risks. To understand how it works and mitigate its effects, our research focuses on scrutinizing the transferability of adversarial attacks in the expanding context of miniature machine learning for UAVs. In this paper, we introduce a formula help measure the transferability of adversarial attacks and explore ways to improve the transferability and effectiveness of adversarial attacks (e.g., a combination of attack techniques), and provide visulizations to vividly illustrate the repercussions of adversarial instances across a spectrum of attack intensities, helping facilitate more intuitive exploration and analysis of the results. For instance, our findings demonstrate that even subtle perturbations directed at specific attributes can lead to a significant decrease in model accuracy. We also evaluates the success rates of various attack algorithms and validates the proposed evaluation methodology for measuring transferability. And the outcomes unveiled in this study make noteworthy strides in fostering a profound comprehension of the transferability of adversarial attacks in the distinct realm of miniature machine learning for UAVs. Robust defense mechanisms, which ensure the impregnability of IoT-enabled UAV systems, can be cultivated by pinpointing the most efficacious attack strategies and evaluating their transferability.