The uncontrollable nature of deployment environments and supply chains for edge devices makes the security of deployed deep neural network (DNN) models a key concern. Leveraging the hardware isolation provided by Trusted Execution Environments (TEEs) to protect sensitive layers of the model is considered a practical solution. However, current studies face the following challenges: (1) The dynamic memory management of mainstream deep learning frameworks is incompatible with the static memory allocation mechanism of TEEs, hindering the integration of an efficient intelligent computing ecosystem throughout the model lifecycle. (2) TEEs lack native support for parallel computing, leading to a performance gap between the TEE and the Rich Execution Environment (REE), which increases overall inference latency. To address these issues, this paper proposes a TEE-based model inference scheme integrated with PyTorch for TrustZone-enabled edge devices. The solution adopts a pre-allocated operator management strategy to eliminate the incompatibility between dynamic graph features and the static memory allocation in the TEE. A semantic-segmentationbased parallel optimization is also introduced to reduce the inference latency of sensitive layers within the TEE. We implement a prototype system on Phytium and evaluate it using four well-known DNN models. Experimental results show that, compared to existing TEE-based DNN inference solutions, our design reduces the lines of code (LoC) for model construction scripts by an average of of 82.2%, lowers secure memory overhead by up to 58.5%, and improves inference performance by an average of$5.38 \times$via octa-core parallel optimization.

Secure execution of smart contracts on consortium blockchains is essential. Existing solutions commonly use trusted execution environments (TEEs) to ensure isolated and confidential contract execution, enhancing security. However, these methods have limitations, such as resource constraints in TEEs, difficulties in supporting various smart contract programming languages, and potential expansion of the TEE’s attack surface. The use of TEE-based integrity measurement for smart contracts offers a viable solution to this problem. However, the challenges of analyzing the scope of integrity attacks must be addressed, determining the object and timing of integrity measurements, and ensuring trust transfer within the contract invocation chain. We study the integrity of consortium blockchain contracts using TEE and establish a smart contract integrity model for the endorsement policy. Furthermore, we propose Tipcc, an integrity measurement framework that combines static system components and dynamic user contracts using TEE. Tipcc ensures integrity measurement and verification throughout the contract lifecycle, providing trusted transmission along the invocation chain. Prototype system validation and simulation experiments using Hyperledger Fabric revealed that this approach improves transaction security while maintaining availability, exhibiting a performance overhead of approximately 7%.

Missing data imputation aims to accurately impute the unobserved regions with complete data in real world. Although many recent methods have made remarkable advances, the local homogenous regions especially in boundary and the reasonable of the imputed data are still two most challenging issues. To address these issues, we propose a novel Global to Local Guiding Network (G2LGN) based on generative adversarial network for missing data imputation, which is composed of a Global-Impute-Net (GIN), a Local-Impute-Net (LIN) and an Impute Guider Model (IGM). The GIN looks at the entire missing regions to generate and impute data as a whole. Considering the reasonable of the GIN results, IGM is assigned to capture coherent information between global and local and guide the LIN to look only at a small area centered at the missing focused regions. After the processing of these three modules, local imputed results are concatenated to global imputed results, which impute reasonable values and refine local details from rough to accurate. The comprehensive experiments on both numeric datasets and image dataset demonstrate our method is significantly superior to other 3 state-of-the-art approaches and 7 traditional methods. Besides, the extensive ablation study validates the superior performance for dealing with missing data imputation.

Generative Adversarial Networks (GANs) have received a lot of attention due to their powerful generative ability, and many related studies have been carried out. Among them, deploying GANs in distributed scenarios has become a hot research topic due to the sharp increase in the capacity of training data. However, distributed GANs face many challenges, such as different data distribution of each node, limitation of cross-node data sampling, etc. The previous methods defaulted to some strong assumptions to alleviate these problems, but the performance would be degraded once they encountered a natural scene. In this paper, we propose Latent Determined Generative Adversarial Network (LDGAN), a network introducing latent determined ensemble to guide the training of the generator without specific assumptions. LDGAN measures the difference in the latent space of data distribution at each node and uses this as the weight of the feedback information of each discriminator, and there is no need for any cross-node data sampling and the independent and identical distribution (iid) data assumption. Our experiments show that on the MNIST, Fashion-MNIST, and CIFAR-10 datasets, the images generated by LDGAN are more realistic and diverse, and the achieved Fréchet Inception Distance (FID) is 15.0%, 22.8%, and 15.3% smaller than the state-of-the-art models, respectively.

Single image reflection removal is a challenging task because important features will be completely obstructed by strong reflections. During removing strong reflections, existing methods focus on mining global contextual features, which will lose some useful local features and fail to remove strong reflections completely. To solve this problem, we propose a feature-integrated patch-hierarchical network (FIPHN), which progressively removes reflections by focusing on integrating global and local features. Specifically, we design a layer-wise feature integration module (LFIM) to integrate global and local features layer by layer across multi-scale stages, effectively enhancing the representation ability of features. Moreover, we design a patch-wise feature integration module (PFIM) to extract contextual features between adjacent patches, avoiding the loss of important features. Meanwhile, with the guidance of ground-truth images, PFIM provides intermediate supervision signals to promote the subnetwork training at each stage. Experimental results show the superiority of our method compared with existing methods.

Defocus blur detection aims to accurately differentiate the defocus blurred regions and in-focus clear regions in a natural image. In spite of remarkable advances, the disturbing of background clutter and the homogenization of boundary remains as the two most challenging issues. To address these issues, we propose an end-to-end network with an Adaptive Dual Attention Module (ADAM), which is designed to simultaneously capture effective local boundary detail and global semantic information in the channel and spatial dimensions, and dynamically aggregate the multi-level features. Specifically, ADAM pays attention to the correlation between local and global features associated with blur objects enhancing the synthetical feature representation. Furthermore, we design a Detail optimization Module (DOM), which is subsequently employed to correct the uncertain detection in boundary regions within the low-resolution details. The comprehensive experiments on two common datasets demonstrate our proposed method has superior performance for dealing with defocus blur detection.

The compression of Generative Adversarial Networks (GANs) has been an emerging study in recent years. However, conventional compression methods can hardly be applied to GANs due to the training process and optimization target of GANs are different from the traditional classification detection network. Inspired by the recent success of knowledge distillation, we condense the recent researches on them applied to particular GANs, i.e. WGAN and CGAN for special tasks into two strategies: Soft Target Only Strategy (STOS) and Inherited Hard Target Strategy (IHTS), and propose an novel Random Hard Target Strategy, namely RHTS. In RHTS, we take the student network generator loss as the hard target of knowledge distillation when the discriminator is fixed, and the output of the teacher network generator as the soft target to distill GANs. The experiments on image datasets (MNIST, CIFAR-10), and structured-data datasets (Australian Credit Approval dataset, Credit Approval Data Set) show that STOS and IHTS are effective, and RHTS is the best. Such results are further proved on WGAN, WGAN-GP and LSGAN in order to demonstrate the better generalization ability of the strategies. Besides, the RHTS improve the stability of the GAN training process, where the performance and stability confirm the rationality of our design.

Small object detection is one of the challenging tasks in recent computer vision researches. The low resolution and noisy representation are internal reasons associated with small object detection. To better address this task, we propose a novel reinforcement learning approach with dynamic refinement. Specifically, we design a Recurrent Reinforcement Module (RRM) to iteratively learn contextual information and improve semantic label dependency and the image-label relevance. Besides, the Dynamic Refinement Module (DRM) is designed to dynamically adjust the attentive regions that are related to small objects. To evaluate the effectiveness of our proposed approach, we design extensive experiments and comparisons on four datasets (i.e. PASCAL VOC, MS COCO, Google AVA and UA-DETRAC). The comprehensive experiment results demonstrate that our approach has the superiority of small object detection for both images and videos.

In recent years, machine learning algorithms and training data are faced many security threats, which affect the security of practical applications based on machine learning. At present, generating adversarial patches based on Generative Adversarial Nets (GANs) has been an emerging study. However, existing attack strategies are still far from producing local adversarial patches with strong attack power, ignoring the attacked network's perceived sensitivity to the adversarial patches. This paper studies the security threat of adversarial patches to classifiers; adding an adversarial patch to the data can mislead the classifier into incorrect results. Considering the attention to aggression and reality, we propose the data-driven MultiD-WGAN, which can simultaneously enhance adversarial patches' attack power and authenticity through multi-discriminators. The experiments confirm that our datadriven MultiD-WGAN dramatically reduces the recall of seven classifiers attacked on four datasets. The attack of data-driven MultiD-WGAN on 25/28 groups of experiments leads to a decreased recall rate, which is better than the conventional GANs. Finally, we have proved a positive correlation between attack intensity and attack ability, both theoretically and experimentally.

Some recent studies have suggested using GANs for numeric data generation such as to generate data for completing the imbalanced numeric data. Considering the significant difference between the dimensions of the numeric data and images, as well as the strong correlations between features of numeric data, the conventional GANs normally face an overfitting problem, consequently leads to an ill-conditioning problem in generating numeric and structured data. This paper studies the constrained network structures between generator G and discriminator D in WGAN, designs several structures including isomorphic, mirror and self-symmetric structures. We evaluates the performances of the constrained WGANs in data augmentations, taking the non-constrained GANs and WGANs as the baselines. Experiments prove the constrained structures have been improved in 16/20 groups of experiments. In twenty experiments on four UCI Machine Learning Repository datasets, Australian Credit Approval data, German Credit data, Pima Indians Diabetes data and SPECT heart data facing five conventional classifiers. Especially, Isomorphic WGAN is the best in 15/20 experiments. Finally, we theoretically proves that the effectiveness of constrained structures by the directed graphic model (DGM) analysis.

In video understanding, a core task is to detect the objects in frames and recently state-of-art methods are proposed to exhaustively detect the possible objects. Few studies are discussing exact these objects by the importance of object roles, which refer to ones that are relatively concerned and attract more attention of the audience in a short video clip. This paper intends to represent the audience's attention transfer mechanism by attractive objects (key object) and proposes a key object detection network based on temporal reinforcement learning (RL). Temporal RL means our model structure using LSTM to output RL reward values dynamically. In a sense, it avoids formal reward setting may cause the RL result does not achieve expectation. In order to mark the object that attracts the audience's attention in successive switching video clips, our method keeps an eye of the attractive object by iterating the value of temporal RL strategy. The proposed model can detect multiple key objects simultaneously with the help of a temporal RL strategy that analyses the attention transfer. Specifically, the spatial features and temporal features, extracted by the ensemble convolution networks, are sent into the RL model to represent the objects, corresponding motions and less relative backgrounds. Google AVA [24] dataset, annotate the position and action of the main characters among many existing objects, are selected as the experiment dataset. Compared with other models, the results show that the proposed method can continuously focus on the attractive objects in the sequenced fragments. Different from the general structure of the deep reinforcement model based on DQN network, our temporal RL parameters are mainly generated and calculated in the KOD-attention block. Therefore, in the process of simulating the attention transfer mechanism, our method uses lightweight computation to achieve satisfactory detection precision and speed in the performance of key object detection.