SoK: Efficiency Robustness of Dynamic Deep Learning Systems

Explore the security vulnerabilities of Dynamic Deep Learning Systems (DDLSs) through this 21-minute conference presentation from USENIX Security '25. Examine how DDLSs, which adapt inference computation based on input complexity to improve efficiency in resource-constrained environments like mobile and IoT devices, introduce new attack surfaces that can be exploited by efficiency adversarial attacks. Learn about the first comprehensive taxonomy of efficiency attacks categorized into three dynamic behaviors: attacks on dynamic computations per inference, attacks on dynamic inference iterations, and attacks on dynamic output production for downstream tasks. Analyze adversarial strategies that specifically target DDLS efficiency and understand the key challenges in securing these adaptive systems. Investigate existing defense mechanisms and discover their limitations against efficiency attacks, highlighting the critical need for novel mitigation strategies to protect future adaptive deep learning deployments in real-time applications.