Efficiency Robustness of Dynamic Deep Learning Systems

Ravishka Rathnasuriya, Zihe Song, Tingxi Li, Zexin Xu, Mirazul Haque, Simin Chen, Wei Yang

Summary

This work examines efficiency attacks against dynamic deep learning systems deployed in resource-constrained settings. It organizes attacks around dynamic computation per inference, dynamic inference iterations, and dynamic output production. It analyzes adversarial strategies and existing defenses, identifies limitations in current mitigation techniques, and highlights challenges in securing adaptive systems.