Exploiting Efficiency Vulnerabilities in Dynamic Deep Learning Systems

The growing deployment of deep learning models in real-world environments has intensified the need for efficient inference under strict latency and resource constraints. To meet these demands, dynamic deep learning systems DDLSs have emerged, offering input-adaptive computation to optimize runtim...