The rapid advancement of pre-trained language models (PLMs) has demonstrated promising results for various code-related tasks. However, their effectiveness in detecting real-world vulnerabilities remains a critical challenge. While existing empirical studies evaluate PLMs for vulnerability detection (VD), they suffer from data leakage, limited scope, and superficial analysis, hindering the accuracy and comprehensiveness of evaluations. This paper begins by revisiting the common issues in existing research on PLMs for VD through the evaluation pipeline. It then proceeds with an accurate and extensive evaluation of 18 PLMs, spanning model parameters from millions to billions, on high-quality datasets that feature accurate labeling, diverse vulnerability types, and various projects. Specifically, we compare the performance of PLMs under both fine-tuning and prompt engineering, assess their effectiveness and generalizability across various training and testing settings, and analyze their robustness to perturbations such as code normalization, abstraction, and semantic-preserving transformations.

Machine learning is evolving towards high-order models that necessitate pre-training on extensive datasets, a process associated with significant overheads. Traditional models, despite having pre-trained weights, are becoming obsolete due to architectural differences that obstruct the effective transfer and initialization of these weights. To address these challenges, we introduce a novel framework, QuadraNet V2, which leverages quadratic neural networks to create efficient and sustainable high-order learning models. Our method initializes the primary term of the quadratic neuron using a standard neural network, while the quadratic term is employed to adaptively enhance the learning of data non-linearity or shifts. This integration of pre-trained primary terms with quadratic terms, which possess advanced modeling capabilities, significantly augments the information characterization capacity of the high-order network. By utilizing existing pre-trained weights, QuadraNet V2 reduces the required GPU hours for training by 90% to 98.4% compared to training from scratch, demonstrating both efficiency and effectiveness.

Federated learning (FL) nowadays involves heterogeneous compound learning tasks as cognitive applications’ complexity increases. For example, a self-driving system hosts multiple tasks simultaneously (e.g., detection, classification, segmentation, etc.) and expects FL to retain life-long intelligence involvement. However, our analysis demonstrates that, when deploying compound FL models for multiple training tasks on a GPU, certain issues arise: As different tasks’ skewed data distributions and corresponding models cause highly imbalanced learning workloads, current GPU scheduling methods lack effective resource allocations; Therefore, existing FL schemes, only focusing on heterogeneous data distribution but runtime computing, cannot practically achieve optimally synchronized federation. To address these issues, we propose a full-stack FL optimization scheme to tackle both intra-device GPU scheduling and inter-device FL coordination for multi-task training. Specifically, our works illustrate two key insights in this research domain: Competitive resource sharing is beneficial for parallel model executions, and the proposed concept of “virtual resource” could effectively characterize and guide the practical per-task resource utilization and allocation; Additionally, architectural-level coordination improves FL performance by aligning task workloads with GPU utilization. Our experiments demonstrate that the FL performance could be significantly escalated. Specifically, we observed a 2.16×–2.38× increase in intra-device GPU training throughput and a 2.53×–2.80× boost in inter-device FL coordination efficiency across diverse multi-task scenarios.

The fusion of multiple modalities, such as vision and language, has led to significant progress in grounding and tracking tasks. However, this success has not yet translated to aerial single-object tracking (SOT) due to the lack of text annotations in existing aerial SOT datasets. To overcome this limitation, we provide text annotations for five existing aerial datasets, designed to support and promote multi-modal research in the aerial tracking domain. Furthermore, to address challenges such as small object dimensions, similar-looking objects, and target size fluctuations, we introduce a third input modality: click (or point prompt). We seamlessly integrate click and language information in the model's input to offer a user-friendly and interactive alternative to precise bounding box annotations. This enables approximate target specification with reduced effort and time. We introduce CLaVi, a novel multimodal framework that redefines input interaction by in-corporating multiple modalities. This integration improves target localization and tracking efficiency, providing a significant advancement in the way input is provided to the model. Furthermore, we conduct experiments on the five datasets, to provide AerTrack-460 benchmark, to validate the effectiveness of our approach. AerTrack-460 benchmark shows competitive performance and, in some cases, outperforms previous language-based grounding and tracking techniques, setting a strong baseline for future research. Code and data will be made available soon.

As the size of Deep Neural Networks (DNNs) continues to grow, their runtime latency also scales. While model pruning and Neural Architecture Search (NAS) can effectively reduce the computation workload, their effectiveness fails to consistently translate into runtime latency reduction. In this paper, we identify the root cause behind the mismatch between workload reduction and latency reduction is GPU tail effect – a classic system issue caused by resource under-utilization in the last processing wave of the GPU. We conduct detailed DNN workload characterization and demonstrate the prevalence of GPU tail effect across different DNN architectures, and meanwhile reveal that the unique deep structure and the light-weight layer workload of DNNs exacerbate the tail effect for DNN inference. We then propose a tail-awareness design space enhancement and DNN optimization algorithm to optimize existing NAS and pruning designs and achieve better runtime latency and model accuracy performance. Extensive experiments show 11%-27% latency reduction over SOTA DNN pruning and NAS methods.

Federated learning (FL), an emerging distributed machine learning paradigm, has been applied to various privacy-preserving scenarios. However, due to its distributed nature, FL faces two key issues: the non-independent and identical distribution (non-IID) of user data and vulnerability to Byzantine threats. To address these challenges, in this paper, we propose FedCAP, a robust FL framework against both data heterogeneity and Byzantine attacks. The core of FedCAP is a model update calibration mechanism to help a server capture the differences in the direction and magnitude of model updates among clients. Furthermore, we design a customized model aggregation rule that facilitates collaborative training among similar clients while accelerating the model deterioration of malicious clients. With a Euclidean norm-based anomaly detection mechanism, the server can quickly identify and permanently remove malicious clients. Moreover, the impact of data heterogeneity and Byzantine attacks can be further mitigated through personalization on the client side. We conduct extensive experiments, comparing multiple state-of-the-art baselines, to demonstrate that FedCAP performs well in several non-IID settings and shows strong robustness under a series of poisoning attacks.

Recent progress in computer vision-oriented neural network designs is mostly driven by capturing high-order neural interactions among inputs and features. And there emerged a variety of approaches to accomplish this, such as Transformers and its variants. However, these interactions generate a large amount of intermediate state and/or strong data dependency, leading to considerable memory consumption and computing cost, and therefore compromising the overall runtime performance. To address this challenge, we rethink the high-order interactive neural network design with a quadratic computing approach. Specifically, we propose QuadraNet — a comprehensive model design methodology from neuron reconstruction to structural block and eventually to the overall neural network implementation. Leveraging quadratic neurons’ intrinsic high-order advantages and dedicated computation optimization schemes, QuadraNet could effectively achieve optimal cognition and computation performance. Incorporating state-of-the-art hardware-aware neural architecture search and system integration techniques, QuadraNet could also be well generalized in different hardware constraint settings and deployment scenarios. The experiment shows that QuadraNet achieves up to 1.5 × throughput, 30% less memory footprint, and similar cognition performance, compared with the state-of-the-art high-order approaches.

As deep learning continues to advance and is applied to increasingly complex scenarios, the demand for concurrent deployment of multiple neural network models has arisen. This demand, commonly referred to as multi-tenant computing, is becoming more and more important. However, even the most mature GPU-based computing systems struggle to adequately address the significant heterogeneity and complexity among concurrent models in terms of resource allocation and runtime scheduling. And this usually results in considerable resource utilization and throughput issues. To tackle these issues, this work proposes a set of optimization techniques that advance the granularity of computing management from both the spatial and temporal perspectives, specifically tailored to heterogeneous model compositions for deep learning inference and training. These techniques are further integrated as GACER --- an automated optimization framework that provides high-utilization, high-throughput, and low-latency multi-tenant computing support. And our experiments demonstrate that GACER significantly improves the overall resource utilization and consistently achieves outstanding speedups compared to native GPU computing frameworks and existing state-of-the-art optimization works.

Recent research works demonstrate that one of the significant factors for the model Out-of-Distirbution detection performance is the scale of the OOD feature representation field. Consequently, model ensemble emerges as a trending method to expand this feature representation field leveraging expected model diversity. However, by proposing novel qualitative and quantitative model ensemble evaluation methods (i.e., Loss Basin/Barrier Visualization and Self-Coupling Index), we reveal that the previous ensemble methods incorporate affine-transformable weights with limited variability and fail to provide desired feature representation diversity. Therefore, we escalate the traditional model ensemble dimensions (different weight initialization, data holdout, etc.) into distinct supervision tasks, which we name as Multi-Comprehension (MC) Ensemble. MC Ensemble leverages various training tasks to form different comprehensions of the data and labels, resulting in the extension of the feature representation field. In experiments, we demonstrate the superior performance of the MC Ensemble strategy in the OOD detection task compared to both the naive Deep Ensemble method and the standalone model of comparable size.

The past neural network design has largely focused on feature \textit{representation space} dimension and its capacity scaling (e.g., width, depth), but overlooked the feature \textit{interaction space} scaling. Recent advancements have shown shifted focus towards element-wise multiplication to facilitate higher-dimensional feature interaction space for better information transformation. Despite this progress, multiplications predominantly capture low-order interactions, thus remaining confined to a finite-dimensional interaction space. To transcend this limitation, classic kernel methods emerge as a promising solution to engage features in an infinite-dimensional space. We introduce InfiNet, a model architecture that enables feature interaction within an infinite-dimensional space created by RBF kernel. Our experiments reveal that InfiNet achieves new state-of-the-art, owing to its capability to leverage infinite-dimensional interactions, significantly enhancing model performance.

Recommendation systems suggest items to users by predicting their preferences based on historical data. The industry traditionally handles large-scale recommendation requests by scaling the number of devices without much concern for a single device’s performance. However, there is a trend for recommendation systems to gradually move from a centralized service to an edge device. The edge-scale recommendation systems have distinct features that are different from traditional large-scale deployments, which poses different challenges to the acceleration of the recommendation system. In this paper, we focus on the edge-scale recommendation system and propose an inter-stage parallelism optimization method deployed on a single GPU. Experiments show that our framework could improve recommendation system throughput by 1.89×~2.2× for different datasets on the GPU.

Convolutional Neural Networks (CNNs) are widely deployed in various embedded recognition applications. However, they demonstrate a considerable vulnerability to adversarial attacks, which leverage the well-designed perturbations to mislead the recognition results. Recently, for easier perturbation injection and higher attack effectiveness, the adversarial perturbations have been concentrated into a small area with various types and different data modalities. When defending such condensed adversarial attacks on the embedded recognition scenarios, most of the existing defense works highlight two critical issues. First, they are particularly designed for each individual condensed attack scenario, lacking enough versatility to accommodate attacks with different data modalities. Second, they rely on computation-intensive preprocessing techniques, which is impractical for time-sensitive embedded recognition scenarios. In this article, we propose LanCeX –a versatile and lightweight CNN defense solution against condensed adversarial attacks. By examining the CNN’s intrinsic vulnerability, we first identify the common attacking mechanism behind condensed adversarial attacks across different data modalities. Based on this mechanism, LanCeX can defend against various condensed attacks with the optimal computation workload in different recognition scenarios. Experiments show that LanCeX can achieve an average 91%, 85%, and 90% detection success rate and optimal adversarial mitigation performance in three recognition scenarios, respectively: image classification, object detection, and audio recognition. Moreover, LanCeX is at most 3× faster compared with the state-of-the-art defense methods, making it feasible to use with resource-constrained embedded systems.

Deep Learning Recommendation Models (DLRM) are widespread, account for a considerable data center footprint, and grow by more than 1.5x per year. With model size soon to be in terabytes range, leveraging Storage Class Memory (SCM) for inference enables lower power consumption. This paper evaluates the major challenges in extending the memory hierarchy to SCM for DLRM, and presents different techniques to improve performance through a Software Defined Memory. We show how underlying technologies such as Nand Flash and 3DXP differentiate, and relate to real world scenarios, enabling from 5% to 29% power savings.

Many works focus on the model’s static parameter optimization (e.g., filters and weights) for CNN inference acceleration. Compared to parameter sparsity, feature map sparsity is per-input related which has better adaptability. The practical sparsity patterns are non-structural and randomly located on feature maps with non-identical shapes. However, the existing feature map sparsity works take computing efficiency as the primary goal, thereby they can only remove structural sparsity and fail to match the above characteristics. In this paper, we develop a novel sparsity computing scheme called FalCon, which can well adapt to the practical sparsity patterns while still maintaining efficient computing. Specifically, we first propose a decomposed convolution design that enables a fine-grained computing unit for sparsity. Additionally, a decomposed convolution computing optimization paradigm is proposed to convert the sparse computing units to practical acceleration. Extensive experiments show that FalCon achieves at most 67.30% theoretical computation reduction with a neglected accuracy drop while accelerating CNN inference by 37%.

Current state-of-the-art object detectors can have significant performance drop when deployed in the wild due to domain gaps with training data. Unsupervised Domain Adaptation (UDA) is a promising approach to adapt detectors for new domains/environments without any expensive label cost. Previous mainstream UDA works for object detection usually focused on image-level and/or feature-level adaptation by using adversarial learning methods. In this work, we show that such adversarial-based methods can only reduce domain style gap, but cannot address the domain content gap that is also important for object detectors. To overcome this limitation, we propose the SC-UDA framework to concurrently reduce both gaps: We propose fine-grained domain style transfer to reduce the style gaps with finer image details preserved for detecting small objects; Then we leverage the pseudo label-based self-training to reduce content gaps; To address pseudo label error accumulation during self-training, novel optimizations are proposed, including uncertainty-based pseudo labeling and imbalanced mini-batch sampling strategy. Experiment results show that our approach consistently outperforms prior state-of-the-art methods (up to 8.6%, 2.7% and 2.5% mAP on three UDA benchmarks).

Personal daily mobility trajectories/traces like Google Location Service integrates many valuable information from individuals and could benefit a lot of application scenarios, such as pandemic control and precaution, product recommendation, customized user profile analysis, traffic management in smart cities, etc. However, utilizing such personal mobility data faces many challenges since users’ private information, such as home/work addresses, can be unintentionally leaked. In this work, we build an FL system for transportation mode prediction based on personal mobility data. Utilizing FL-based training scheme, all user’s data are kept in local without uploading to central nodes, providing high privacy preserving capability. At the same time, we could train accurate DNN models that is close to the centralized training performance. The resulted transportation mode prediction system serves as a prototype on user’s traffic mode classification, which could potentially benefit the transportation data analysis and help make wise decisions to manage public transportation resources.

Nowadays, the evolution of deep learning and cloud service significantly promotes neural network-based mobile applications. Although intelligent and prolific, those applications still lack certain flexibility: for classification tasks, neural networks are generally trained with vast classification targets to cover various utilization contexts. However, only partial classes are practically inferred due to individual mobile user preference and application specificity, which causes unnecessary computation consumption. Thus, we proposedCaptorX—a class-adaptive convolutional neural network (CNN) reconfiguration framework to adaptively prune convolutional filters associated with unneeded classes.CaptorXcan reconfigure a pretrained full-class CNN model into class-specific lightweight models based on the visualization analysis of convolutional filters’ exclusive functionality for a single class. These lightweight models can be directly deployed to mobile devices without the retraining cost of traditional pruning-based reconfiguration. Furthermore, we can apply theCaptorXframework into a distributed collaboration setting. With dedicated local training regulation and collaborative aggregation schemes, the class-adaptive models on individual mobile devices can further contribute back to the central full-class model. Experiments on representative CNNs and image classification datasets show that,CaptorXcan reduce the CNN computation workload up to 50.22% and save 46.58% energy consumption for varied local devices, meanwhile improving accuracy for their targeted classes with better task focus. With our distributed collaboration paradigm,CaptorXalso provides further potential to enhance the central model accuracy, while reducing up to 37.58% communication cost compared to traditional distributed learning methods.

Deep neural networks (DNNs) achieved great cognitive performance at the expense of a considerable computation workload. To relieve the computational burden, many optimization works are developed to reduce the model redundancy by identifying and removing insignificant model components, such as weight sparsity and filter pruning methods. However, these works only evaluate model components’ static significance with parameter information, ignoring their dynamic interaction with external inputs. Specifically, due to the difference in per-input features, the model components’ significance can dynamically change and, thus, the static methods can only achieve suboptimal performance. Focusing on this aspect, we propose a dynamic DNN optimization framework in this work. Based on the neural network attention mechanism, we propose a comprehensive dynamic optimization framework, including 1) testing-phase dynamic feature map pruning; 2) training-phase optimization by training with targeted dropout; and 3) deployment-phase one-for-all (OFA) model adaptability enhancement. By providing a holistic dynamic testing, training, and deployment co-optimization framework, our work has the following benefits: first, it can accurately identify and aggressively remove per-input feature redundancy by considering the model-input interaction and involving the channel/column-wise pruning flexibility; meanwhile, the training-testing co-optimization favors the dynamic pruning and helps maintain the model accuracy even with a very high feature pruning ratio. Finally, the deployment enhancement provides one unified OFA model to support full-spectrum feature sparsity ratios. The unified model can be dynamically reconfigured to meet different resource budgets without any retraining cost, and thus provide significant deployment flexibility. Extensive experiments show that our method could bring 37.4%–54.5% floating-point operations reduction with negligible accuracy drop on various test benchmarks. Meanwhile, the OFA deployment optimization enables us to use one model to support at most ten different resource constraints without any retraining cost.

As Federated Learning (FL) has been widely used for collaborative training, a considerable computational straggler issue emerged: when FL deploys identical neural network models to heterogeneous devices, the ones with weak computational capacities, referred to as stragglers, may significantly delay the synchronous parameter aggregation. Although discarding stragglers from the collaboration can relieve this issue to a certain extent, stragglers may keep unique and critical information learned from the non-identical dataset, and directly discarding will harm the overall collaboration performance. Therefore, in this paper, we propose Helios – a heterogeneity-aware FL framework to tackle the straggler issue. Helios identifies individual devices’ heterogeneous training capability, and therefore the expected neural network model training volumes regarding the collaborative training pace. For straggling devices, a “softtraining” method is proposed to dynamically compress the original identical training model into the expected volume through a rotated neuron training approach. With extensive algorithm analysis and optimization schemes, stragglers can be accelerated while retaining the convergence for local training as well as federated collaboration. Experiments show that Helios can provide up to $2.5\times$ training acceleration and maximum 4.64% convergence accuracy improvement in various collaboration settings.

With the fast development of deep neural networks (DNNs), many real-world applications are adopting multiple models to conduct compound tasks, such as co-running classification, detection, and segmentation models on autonomous vehicles. Such multi-tenant DNN inference cases greatly exacerbate the computational complexity and call for comprehensive collaboration for graph-level operator scheduling, runtime-level resource awareness, as well as hardware scheduler support. However, the current scheduling support for such multi-tenant inference is still relatively backward. In this work, we propose a resource-aware scheduling framework for efficient multi-tenant DNN inference on GPU, which automatically coordinates DNN computing in different execution levels. Leveraging the unified scheduling intermediate representation and the automated ML-based searching algorithm, optimal schedules could be generated to wisely adjust model concurrency and interleave DNN model operators, maintaining a continuously balanced resource utilization across the entire inference process, and eventually improving the runtime efficiency. Experiments show that we could consistently achieve$1.3\times\sim 1.7\times$speed-up, comparing to regular DNN runtime libraries (e.g., CuDNN, TVM) and particular concurrent scheduling methods (e.g., NVIDIA Multi-Stream).

Federated learning learns from scattered data by fusing collaborative models from local nodes. However, conventional coordinate-based model averaging by FedAvg ignored the random information encoded per parameter and may suffer from structural feature misalignment. In this work, we propose Fed2, a feature-aligned federated learning framework to resolve this issue by establishing a firm structure-feature alignment across the collaborative models. Fed2 is composed of two major designs: First, we design a feature-oriented model structure adaptation method to ensure explicit feature allocation in different neural network structures. Applying the structure adaptation to collaborative models, matchable structures with similar feature information can be initialized at the very early training stage. During the federated learning process, we then propose a feature paired averaging scheme to guarantee aligned feature distribution and maintain no feature fusion conflicts under either IID or non-IID scenarios. Eventually, Fed2 could effectively enhance the federated learning convergence performance under extensive homo- and heterogeneous settings, providing excellent convergence speed, accuracy, and computation/communication efficiency.

Although convolutional neural networks (CNNs) have been widely applied in various cognitive applications, they are still very computationally intensive for resource-constrained mobile systems. To reduce the resource consumption of CNN computation, many optimization works have been proposed for mobile CNN deployment. However, most works are merely targeting CNN model compression from the perspective of parameter size or model structure, ignoring different resource constraints in mobile systems with respect to memory, energy, and real-time requirement. Moreover, previous works take accuracy as their primary consideration, requiring a time-costing retraining process to compensate the inference accuracy loss after compression. To address these issues, we propose DiReCtX-a dynamic resource-aware CNN model reconfiguration framework. DiReCtX is based on a set of accurate CNN profiling models for different resource consumption and inference accuracy estimation. With manageable consumption/accuracy tradeoffs, DiReCtX can reconfigure a CNN model to meet distinct resource constraint types and levels with expected inference performance maintained. To further achieve fast model reconfiguration in real-time, improved CNN model pruning and its corresponding accuracy tuning strategies are also proposed in DiReCtX. The experiments show that the proposed CNN profiling models can achieve 94.6% and 97.1% accuracy for CNN model resource consumption and inference accuracy estimation. Meanwhile, the proposed reconfiguration scheme of DiReCtX can achieve at most 44.44% computation acceleration, 31.69% memory reduction, and 32.39% energy saving, respectively. On field-tests with state-of-the-art smartphones, DiReCtX can adapt CNN models to various resource constraints in mobile application scenarios with optimal real-time performance.

In recent years, the neural network (NN) has shown its great potential in image recognition tasks of autonomous driving systems, such as traffic sign recognition, pedestrian detection, etc. However, theoretically well-trained NNs usually fail their performance when facing real-world scenarios. For example, adverse real-world conditions, e.g., bad weather and lighting conditions, can introduce different physical variations and cause considerable accuracy degradation. As for now, the generalization capability of NNs is still one of the most critical challenges for the autonomous driving system. To facilitate the robust image recognition tasks, in this work, we build the RobuTS dataset: a comprehensive Robust Traffic Sign Recognition dataset, which includes images with different environmental variations, e.g., rain, fog, darkening, and blurring. Then to enhance the NN's generalization capability, we propose two generalization-enhanced training schemes: 1) REIN for robust training without data in adverse scenarios and 2) Self-Teaching (ST) for robust training with unlabeled adverse data. The great advantages of such two training schemes are they are data-free (REIN) and label-free (ST), thus effectively reducing the huge human efforts/cost of on-road driving data collection, as well as the expensive manual data annotation. We conduct extensive experiments to validate our methods' performance on both classification and detection tasks. For classification tasks, our proposed training algorithms could consistently improve model performance by +15%-25% (REIN) and +16%-30% (ST) in all adverse scenarios of our RobuTS datasets. For detection tasks, our ST could also improve the detector's performance by +10.1 mean average precision (mAP) on Foggy-Cityscapes, outperforming previous state-of-the-art works by +2.2 mAP.

The memristor crossbar array has emerged as an intrinsically suitable matrix computation and low-power acceleration framework for DNN applications. Many techniques such as memristor-based weight pruning and memristor-based quantization have been studied. However, the high accuracy solution for the above techniques is still waiting for unraveling. In this paper, we propose a memristor-based DNN framework which combines both structured weight pruning and quantization by incorporating ADMM algorithm for better pruning and quantization performance. We also discover the non-optimality of the ADMM solution in weight pruning and the unused data path in a structured pruned model. We design a software-hardware co-optimization framework which contains the first proposed Network Purification and Unused Path Removal algorithms targeting on post-processing a structured pruned model after ADMM steps. By taking memristor hardware constraints into our whole framework, we achieve extreme high compression rate with minimum accuracy loss. For quantizing structured pruned model, our framework achieves nearly no accuracy loss after quantizing weights to 8-bit memristor weight representation. We share our models at anonymous link https://bit.ly/2VnMUy0.

Recently, adversarial attacks can be applied to the physical world, causing practical issues to various Convolutional Neural Networks (CNNs) powered applications. Most existing physical adversarial attack defense works only focus on eliminating explicit perturbation patterns from inputs, ignoring interpretation to CNN's intrinsic vulnerability. Therefore, they lack expected versatility to different attacks and thereby depend on considerable data processing costs. In this paper, we propose LanCe - a comprehensive and lightweight CNN defense methodology against different physical adversarial attacks. By interpreting CNN's vulnerability, we find that non-semantic adversarial perturbations can activate CNN with significantly abnormal activations and even overwhelm other semantic input patterns' activations. We improve the CNN recognition process by adding a self-verification stage to detect the potential adversarial input with only one CNN inference cost. Based on the detection result, we further propose a data recovery methodology to defend the physical adversarial attacks. We apply such defense methodology into both image and audio CNN recognition scenarios and analyze the computational complexity for each scenario, respectively. Experiments show that our methodology can achieve an average 91% successful rate for attack detection and 89% accuracy recovery. Moreover, it is at most 3× faster compared with the state-of-the-art defense methods, making it feasible to resource-constrained embedded systems, such as mobile devices.

Convolutional Neural Networks (CNNs) achieved great cognitive performance at the expense of considerable computation load. To relieve the computation load, many optimization works are developed to reduce the model redundancy by identifying and removing insignificant model components, such as weight sparsity and filter pruning. However, these works only evaluate model components’ static significance with internal parameter information, ignoring their dynamic interaction with external inputs. With per-input feature activation, the model component significance can dynamically change, and thus the static methods can only achieve sub-optimal results. Therefore, we propose a dynamic CNN optimization framework in this work. Based on the neural network attention mechanism, we propose a comprehensive dynamic optimization framework including (1) testing-phase channel and column feature map pruning, as well as (2) training-phase optimization by targeted dropout. Such a dynamic optimization framework has several benefits: (1) First, it can accurately identify and aggressively remove per-input feature redundancy with considering the model-input interaction; (2) Meanwhile, it can maximally remove the feature map redundancy in various dimensions thanks to the multi-dimension flexibility; (3) The training-testing co-optimization favors the dynamic pruning and helps maintain the model accuracy even with very high feature pruning ratio. Extensive experiments show that our method could bring 37.4%∼54.5% FLOPs reduction with negligible accuracy drop on various of test networks.

Recently Convolutional Neural Networks (CNNs) are widely applied into novel intelligent applications and systems. However, the CNN computation performance is significantly hindered by its computation flow, which computes the model structure sequentially by layers with massive convolution operations. Such a layer-wise sequential computation flow can cause certain performance issues, such as resource under-utilization, huge memory overhead, etc. To solve these problems, we propose a novel CNN structural decoupling method, which could decouple CNN models into "critical paths" and eliminate the inter-layer data dependency. Based on this method, we redefine the CNN computation flow into parallel and cascade computing paradigms, which can significantly enhance the CNN computation performance with both multi-core and single-core CPU processors. Experiments show that, our DC-CNN framework could reduce 24% to 33% latency on multi-core CPUs for CIFAR and ImageNet. On small-capacity mobile platforms, cascade computing could reduce the latency by average 24% on ImageNet and 42% on CIFAR10. Meanwhile, the memory reduction could also reach average 21% and 64%, respectively.

This paper gives an overview of our ongoing work on the design space exploration of efficient deep neural networks (DNNs), specifically on the novel optimization perspectives that past work have mainly overlooked. We cover two complementary aspects of efficient DNN design: (1) static architecture design efficiency and (2) dynamic model execution efficiency. In the static architecture design, one of the major challenges of NAS is the low search efficiency. Different with current mainstream efficient search algorithm optimization, we identify the new perspective in efficient search space design. In the dynamic model execution, current major optimization methods still target at the model structure redundancy, e.g., weight/filter pruning, connection pruning, etc. We instead identify the new dimension of DNN feature map redundancy. By showcasing such new perspectives, further advantages could be potentially attained by integrating both current optimizations and our new perspectives.

Resistive random-access memory (ReRAM) based accelerators have been widely studied to achieve efficient neural network computing in speed and energy. Neural network optimization algorithms such as sparsity are developed to achieve efficient neural network computing on traditional computer architectures such as CPU and GPU. However, such computing efficiency improvement is hindered when deploying these algorithms on the ReRAM-based accelerator because of its unique crossbar-structural computations. And a specific algorithm and hardware co-optimization for the ReRAM-based architecture is still in a lack. In this work, we propose an efficient neural network computing framework that is specialized for the crossbar-structural computations on the ReRAM-based accelerators. The proposed framework includes a crossbar specific feature map pruning and an adaptive neural network deployment. Experimental results show our design can improve the computing accuracy by 9.1% compared with the state-of-the-art sparse neural networks. Based on a famous ReRAM-based DNN accelerator, the proposed framework demonstrates up to 1.4× speedup, 4.3× power efficiency, and 4.4× area saving.

Nowadays, the evolution of deep learning and cloud service significantly promotes neural network based mobile applications. Although intelligent and prolific, those applications still lack certain flexibility: For classification tasks, neural networks are generally trained online with vast classification targets to cover various utilization contexts. However, only partial classes are practically tested due to individual mobile user preference and application specificity. Thus the unneeded classes cause considerable computation and communication cost. In this work, we propose CAPTOR - a class-level reconfiguration framework for Convolutional Neural Networks (CNNs). By identifying the class activation preference of convolutional filters through feature interest visualization and gradient analysis, CAPTOR can effectively cluster and adaptively prune the filters associated with unneeded classes. Therefore, CAPTOR enables class-level CNN reconfiguration for network model compression and local deployment on mobile devices. Experiment shows that, CAPTOR can reduce computation load for VGG-16 by up to 40.5% and 37.9% energy consumption with ignored loss of accuracy. For AlexNet, CAPTOR also reduces computation load by up to 42.8% and 37.6% energy consumption with less than 3% loss in accuracy.

Recently, the machine learning technologies have been widely used in cognitive applications such as Automatic Speech Recognition (ASR) and Image Recognition (IR). Unfortunately, these techniques have been massively used in unauthorized audio/image data analysis, causing serious privacy leakage. To address this issue, we propose HAMPER in this work, which is a data encryption framework that protects the audio/image data from unauthorized ASR/IR analysis. Leveraging machine learning models' vulnerability to adversarial examples, HAMPER encrypt the audio/image data with adversarial noises to perturb the recognition results of ASR/IR systems. To deploy the proposed framework in extensive platforms (e.g. mobile devices), HAMPER also take into consideration of computation efficiency, perturbation transferability, as well as data attribute configuration. Therefore, rather than focusing on the high-level machine learning models, HAMPER generates adversarial examples from the low-level features. Taking advantage of the light computation load, fundamental impact, and direct configurability of the low-level features, the generated adversarial examples can efficiently and effectively affect the whole ASR/IR systems. Experiment results show that, HAMPER can effectively perturb the unauthorized ASR/IR analysis with 85% Word-Error-Rate (WER) and 83% Image-Error-Rate (IER) respectively. Also, HAMPER achieves faster processing speed with 1.5X speedup for image encryption and even 26X in audio, comparing to the state-of-the-art methods. Moreover, HAMPER achieves strong transferability and configures adversarial examples with desired attributes for better scenario adaptation.

In recent years, neural network has shown its great potential in autonomous driving systems. However, the theoretically well-train neural networks usually fail their performance when facing real-world examples with unexpected physical variations. As the current neural networks still suffer from limited generalization ability, those unexpected variations would cause considerable accuracy degradation and critical safety issues. Therefore, the generalization ability of neural networks becomes one of the most critical challenges for autonomous driving system design. In this work, we propose a robust training method to enhance neural network's generalization ability in various practical autonomous driving scenarios. Based on detailed practical variation modeling and neural network generation ability analysis, the proposed training method could consistently improve model classification accuracy by at most 25% in various scenarios (e.g. raining/fogy, dark lighting, and camera discrepancy). Even with adversarial corner cases, our model could still achieve at most 40% accuracy improvement over natural model.

Although the Deep Neural Network (DNN) technique has been widely applied in various applications, the DNN-based applications are still too computationally intensive for the resource-constrained mobile devices. Many works have been proposed to optimize the DNN computation performance, but most of them are limited in an algorithmic perspective, ignoring certain computing issues in practical deployment. To achieve the comprehensive DNN performance enhancement in practice, the expected DNN optimization works should closely cooperate with specific hardware and system constraints (i.e. computation capacity, energy cost, memory occupancy, and inference latency). Therefore, in this work, we propose ReForm -- a resource-aware DNN optimization framework. Through thorough mobile DNN computing analysis and innovative model reconfiguration schemes (i.e. ADMM based static model fine-tuning, dynamically selective computing), ReForm can efficiently and effectively reconfigure a pre-trained DNN model for practical mobile deployment with regards to various static and dynamic computation resource constraints. Experiments show that ReForm has ~3.5× faster optimization speed than state-of-the-art resource-aware optimization method. Also, ReForm can effective reconfigure a DNN model to different mobile devices with distinct resource constraints. Moreover, ReForm achieves satisfying computation cost reduction with ignorable accuracy drop in both static and dynamic computing scenarios (at most 18% workload, 16.23% latency, 48.63% memory, and 21.5% energy enhancement).

Benefited from recent artificial intelligence evolution, Automatic Speech Recognition (ASR) technology has achieved enormous performance improvement and wider application. Unfortunately, ASR is also heavily leveraged by speech eavesdropping, where ASR is used to translate large volume of intercepted vocal speech into text content, causing considerable information leakage. In this work, we propose MASKER -- a mobile security enhancement solution to protect the mobile speech data from ASR in eavesdropping. By identifying ASR models' ubiquitous vulnerability, MASKER is designed to generate human imperceptible adversarial noises into the real-time speech on the mobile device (e.g. phone call and voice message). Even the speech data is exposed to eavesdropping during data transmission, the adversarial noises can effectively perturb the ASR process with significant Word Error Rate (WER). Meanwhile, MASKER is further optimized for mobile user perception quality and enhanced for environmental noises adaptation. Moreover, MASKER has outstanding computation efficiency for mobile system integration. Experiments show that, MASKER can achieve security enhancement with an average WER of 84.55% for ASR perturbation, 32% noise reduction for user perception quality and 16× faster processing speed compared to the state-of-the-art method.

Recently, adversarial deception becomes one of the most considerable threats to deep neural networks. However, compared to extensive research in new designs of various adversarial attacks and defenses, the neural networks' intrinsic robustness property is still lack of thorough investigation. This work aims to qualitatively interpret the adversarial attack and defense mechanisms through loss visualization, and establish a quantitative metric to evaluate the model's intrinsic robustness. The proposed robustness metric identifies the upper bound of a model's prediction divergence in the given domain and thus indicates whether the model can maintain a stable prediction. With extensive experiments, our metric demonstrates several advantages over conventional testing accuracy based robustness estimation: (1) it provides a uniformed evaluation to models with different structures and parameter scales; (2) it over-performs conventional accuracy based robustness evaluation and provides a more reliable evaluation that is invariant to different test settings; (3) it can be fast generated without considerable testing cost.

Alternating Direction Method of Multipliers (ADMM) has been used successfully in many conventional machine learning applications and is considered to be a useful alternative to Stochastic Gradient Descent (SGD) as a deep learning optimizer. However, as an emerging domain, several challenges remain, including 1) The lack of global convergence guarantees, 2) Slow convergence towards solutions, and 3) Cubic time complexity with regard to feature dimensions. In this paper, we propose a novel optimization framework for deep learning via ADMM (dlADMM) to address these challenges simultaneously. The parameters in each layer are updated backward and then forward so that the parameter information in each layer is exchanged efficiently. The time complexity is reduced from cubic to quadratic in (latent) feature dimensions via a dedicated algorithm design for subproblems that enhances them utilizing iterative quadratic approximations and backtracking. Finally, we provide the first proof of global convergence for an ADMM-based method (dlADMM) in a deep neural network problem under mild conditions. Experiments on benchmark datasets demonstrated that our proposed dlADMM algorithm outperforms most of the comparison methods.

Although Convolutional Neural Networks (CNNs) have been widely applied in various applications, their deployment in resource-constrained mobile systems remains a significant concern. To overcome the computation resource constraints, such as limited memory and energy capacity, many works are proposed for mobile CNN optimization. However, most of them lack a comprehensive modeling analysis of the CNN computation consumption and merely focus on static optimization schemes regardless of different mobile computation scenarios. In this work, we proposed DiReCt -- a resource-aware CNN reconfiguration system. Leveraging accurate CNN computation consumption modeling and mobile resource constraint analysis, DiReCt can reconfigure a CNN with different accuracy and resource consumption levels to adapt to various mobile computation scenarios. The experiment results show that: the proposed computation consumption models in DiReCt can well estimate the CNN computation consumption with 94.1% accuracy, and DiReCt achieves at most 34.9% computation acceleration, 52.7% memory reduction, and 27.1% energy saving. Eventually, DiReCt can effectively adapt CNNs to dynamic mobile usage scenarios for optimal performance.