Neural image caption generation (NICG) models have received massive attention from the research community due to their excellent performance in visual understanding. Existing work focuses on improving NICG model ac-curacy while efficiency is less explored. However, many real-world applications require real-time feedback, which highly relies of models. Recent re-search observed that could vary for different inputs. This observation brings a new attack surface models, i.e., An adversary might...

Neural Machine Translation (NMT) systems have received much recent attention due to their human-level accuracy. While existing works mostly focus on either improving accuracy or testing robustness, the computation efficiency of NMT systems, which is paramount importance often vast translation demands and real-time requirements, has surprisingly little attention. In this paper, we make first attempt understand test potential robustness in state-of-the-art systems. By analyzing working...

Today, an increasing number of Adaptive Deep Neural Networks (AdNNs) are being used on resource-constrained embedded devices. We observe that, similar to traditional software, redundant computation exists in AdNNs, resulting considerable performance degradation. The degradation is dependent the input and referred as input-dependent bottlenecks (IDPBs). To ensure AdNN satisfies requirements applications, it essential conduct testing detect IDPBs AdNN. Existing neural network methods primarily...

With the increasing number of layers and parameters in neural networks, energy consumption networks has become a great concern to society, especially users handheld or embedded devices. In this paper, we investigate robustness against energy-oriented attacks. Specifically, propose ILFO (Intermediate Output-Based Loss Function Optimization) attack common type energy-saving Adaptive Neural Networks (AdNN). AdNNs save by dynamically deactivating part its model based on need inputs. leverages...

Recent advancements in deploying deep neural networks (DNNs) on resource-constrained devices have generated interest input-adaptive dynamic (DyNNs). DyNNs offer more efficient inferences and enable the deployment of DNNs with limited resources, such as mobile devices. However, we discovered a new vulnerability that could potentially compromise their efficiency. Specifically, investigate whether adversaries can manipulate DyNNs' computational costs to create false sense To address this...

Recently, various Deep Neural Network (DNN) models have been proposed for environments like embedded systems with stringent energy constraints. The fundamental problem of determining the robustness a DNN respect to its consumption (energy robustness) is relatively unexplored compared accuracy-based robustness. This work investigates Adaptive Networks (AdNNs), type energy-saving DNNs many energy-sensitive domains and recently gained traction. We propose EREBA, first black-box testing method...

The recently proposed capability-based NLP testing allows model developers to test the functional capabilities of models, revealing failures that cannot be detected by traditional heldout mechanism. However, existing work on requires extensive manual efforts and domain expertise in creating cases. In this paper, we investigate a low-cost approach for case generation leveraging GPT-3 engine. We further propose use classifier remove invalid outputs from expand into templates generate more Our...

Recently, Neural ODE (Ordinary Differential Equation) models have been proposed, which use ordinary differential equation solving to predict the output of neural networks. Due models' noticeably lower parameter usage compared traditional Deep Networks (DNN) and higher robustness against gradient-based attacks, they are being adopted in many type real-time applications. For applications, response-time (latency) has paramount importance due convenience user. Through our observation, we find...

Deep Neural Networks (DNNs) have been used to solve different day-to-day problems. Recently, DNNs deployed in real-time systems, and lowering the energy consumption response time has become need of hour. To address this scenario, researchers proposed incorporating dynamic mechanism static (SDNN) create Dynamic (DyNNs) performing amounts computation based on input complexity. Although into SDNNs would be preferable it also becomes important evaluate how introduction impacts robustness models....

Because of the increasing accuracy Deep Neural Networks (DNNs) on different tasks, a lot real times systems are utilizing DNNs. These DNNs vulnerable to adversarial perturbations and corruptions. Specifically, natural corruptions like fog, blur, contrast etc can affect prediction DNN in an autonomous vehicle. In time, these needed be detected also corrupted inputs denoised predicted correctly. this work, we propose CorrGAN approach, which generate benign input when is provided. framework,...

Deep Learning (DL) models have been popular nowadays to execute different speech-related tasks, including automatic speech recognition (ASR). As ASR is being used in real-time scenarios, it important that the model remains efficient against minor perturbations input. Hence, evaluating efficiency robustness of need hour. We show like Speech2Text and Whisper dynamic computation based on inputs, causing efficiency. In this work, we propose SlothSpeech, a denial-of-service attack models, which...

Deep Neural Networks (DNNs) have been used to solve different day-to-day problems. Recently, DNNs deployed in real-time systems, and lowering the energy consumption response time has become need of hour. To address this scenario, researchers proposed incorporating dynamic mechanism static (SDNN) create Dynamic (DyNNs) performing amounts computation based on input complexity. Although into SDNNs would be preferable it also becomes important evaluate how introduction impacts robustness models....

Neural image caption generation (NICG) models have received massive attention from the research community due to their excellent performance in visual understanding. Existing work focuses on improving NICG model accuracy while efficiency is less explored. However, many real-world applications require real-time feedback, which highly relies of models. Recent observed that could vary for different inputs. This observation brings a new attack surface models, i.e., An adversary might be able...

Because of the increasing accuracy Deep Neural Networks (DNNs) on different tasks, a lot real times systems are utilizing DNNs. These DNNs vulnerable to adversarial perturbations and corruptions. Specifically, natural corruptions like fog, blur, contrast etc can affect prediction DNN in an autonomous vehicle. In time, these needed be detected also corrupted inputs de-noised predicted correctly. this work, we propose CorrGAN approach, which generate benign input when is provided. framework,...

Neural Machine Translation (NMT) systems have received much recent attention due to their human-level accuracy. While existing works mostly focus on either improving accuracy or testing robustness, the computation efficiency of NMT systems, which is paramount importance often vast translation demands and real-time requirements, has surprisingly little attention. In this paper, we make first attempt understand test potential robustness in state-of-the-art systems. By analyzing working...

Today, an increasing number of Adaptive Deep Neural Networks (AdNNs) are being used on resource-constrained embedded devices. We observe that, similar to traditional software, redundant computation exists in AdNNs, resulting considerable performance degradation. The degradation is dependent the input and referred as input-dependent bottlenecks (IDPBs). To ensure AdNN satisfies requirements applications, it essential conduct testing detect IDPBs AdNN. Existing neural network methods primarily...