Efficiency Robustness of Dynamic Deep Learning Systems

Deep Learning Systems (DLSs) are increasingly deployed in real-time applications, including those in resourceconstrained environments such as mobile and IoT devices. To address efficiency challenges, Dynamic Deep Learning Systems (DDLSs) adapt inference computation based on input complexity, reducing overhead. While this dynamic behavior improves efficiency, such behavior introduces new attack surfaces. In particular, efficiency adversarial attacks exploit these dynamic mechanisms to degrade system performance. This paper systematically explores efficiency robustness of DDLSs, presenting the first comprehensive taxonomy of efficiency attacks. We categorize these attacks based on three dynamic behaviors: (i) attacks on dynamic computations per inference, (ii) attacks on dynamic inference iterations, and (iii) attacks on dynamic output production for downstream tasks. Through an in-depth evaluation, we analyze adversarial strategies that target DDLSs efficiency and identify key challenges in securing these systems. In addition, we investigate existing defense mechanisms, demonstrating their limitations against increasingly popular efficiency attacks and the necessity for novel mitigation strategies to secure future adaptive DDLSs.

Testing Hateful Speeches against Policies

In the recent years, many software systems have adopted AI techniques, especially deep learning techniques. Due to their black-box nature, AI-based systems brought challenges to traceability, because AI system behaviors are based on models and data, whereas the requirements or policies are rules in the form of natural or programming language. To the best of our knowledge, there is a limited amount of studies on how AI and deep neural network-based systems behave against rule-based requirements/policies. This experience paper examines deep neural network behaviors against rule-based requirements described in natural language policies. In particular, we focus on a case study to check AI-based content moderation software against content moderation policies. First, using crowdsourcing, we collect natural language test cases which match each moderation policy, we name this dataset HateModerate; second, using the test cases in HateModerate, we test the failure rates of state-of-the-art hate speech detection software, and we find that these models have high failure rates for certain policies; finally, since manual labeling is costly, we further proposed an automated approach to augument HateModerate by finetuning OpenAI's large language models to automatically match new examples to policies. The dataset and code of this work can be found on our anonymous website: \url{https://sites.google.com/view/content-moderation-project}.