<ul data-eligibleForWebStory="true"><li>Deep Learning Systems are being widely used in real-time applications, including resource-constrained environments like mobile and IoT devices.</li><li>Dynamic Deep Learning Systems (DDLSs) adjust inference computation based on input complexity to improve efficiency and reduce overhead.</li><li>However, the dynamic nature of DDLSs opens up vulnerabilities to efficiency adversarial attacks, which exploit these adaptive mechanisms to degrade system performance.</li><li>This paper introduces a taxonomy of efficiency attacks on DDLSs, categorizing them into three types based on dynamic behaviors.</li><li>The identified attack categories include those focusing on dynamic computations per inference, dynamic inference iterations, and dynamic output production for downstream tasks.</li><li>The study delves into adversarial strategies targeting DDLSs' efficiency and highlights the challenges in securing these systems.</li><li>Existing defense mechanisms are examined, revealing their limitations against the evolving landscape of efficiency attacks, urging the need for innovative mitigation strategies.</li></ul>

Efficiency Robustness of Dynamic Deep Learning Systems

Discover more