Biblio

Deep Neural Networks (DNNs) have been widely used in variety of fields with great success. However, recent researches indicate that DNNs are susceptible to adversarial attacks, which can easily fool the well-trained DNNs without being detected by human eyes. In this paper, we propose to combine the target DNN model with robust bit plane classifiers to defend against adversarial attacks. It comes from our finding that successful attacks generate imperceptible perturbations, which mainly affects the low-order bits of pixel value in clean images. Hence, using bit planes instead of traditional RGB channels for convolution can effectively reduce channel modification rate. We conduct experiments on dataset CIFAR-10 and GTSRB. The results show that our defense method can effectively increase the model accuracy on average from 8.72% to 85.99% under attacks on CIFAR-10 without sacrificina accuracy of clean images.

Recent advances in adversarial attacks uncover the intrinsic vulnerability of modern deep neural networks. Since then, extensive efforts have been devoted to enhancing the robustness of deep networks via specialized learning algorithms and loss functions. In this work, we take an architectural perspective and investigate the patterns of network architectures that are resilient to adversarial attacks. To obtain the large number of networks needed for this study, we adopt one-shot neural architecture search, training a large network for once and then finetuning the sub-networks sampled therefrom. The sampled architectures together with the accuracies they achieve provide a rich basis for our study. Our ''robust architecture Odyssey'' reveals several valuable observations: 1) densely connected patterns result in improved robustness; 2) under computational budget, adding convolution operations to direct connection edge is effective; 3) flow of solution procedure (FSP) matrix is a good indicator of network robustness. Based on these observations, we discover a family of robust architectures (RobNets). On various datasets, including CIFAR, SVHN, Tiny-ImageNet, and ImageNet, RobNets exhibit superior robustness performance to other widely used architectures. Notably, RobNets substantially improve the robust accuracy ( 5% absolute gains) under both white-box and black-box attacks, even with fewer parameter numbers. Code is available at https://github.com/gmh14/RobNets.

Battery energy storage systems are facing risks of unreliable battery sensor data which might be caused by sensor faults in an embedded battery management system, communication failures, and even cyber-attacks. It is crucial to evaluate the trustworthiness of battery sensor data since inaccurate sensor data could lead to not only serious damages to battery energy storage systems, but also threaten the overall reliability of their applications (e.g., electric vehicles or power grids). This paper introduces a battery sensor data trust framework enabling detecting unreliable data using a deep learning algorithm. The proposed sensor data trust mechanism could potentially improve safety and reliability of the battery energy storage systems. The proposed deep learning-based battery sensor fault detection algorithm is validated by simulation studies using a convolutional neural network.

Nowadays, the networks attacks became very sophisticated and hard to be recognized, The traditional types of intrusion detection systems became inefficient in predicting new types of attacks. As the IDS is an important factor in securing the network in the real time, many new effective IDS approaches have been proposed. In this paper, we intend to discuss different Artificial Neural Networks based IDS approaches, also we are going to categorize them in four categories (normal ANN, DNN, CNN, RNN) and make a comparison between them depending on different performance parameters (accuracy, FNR, FPR, training time, epochs and the learning rate) and other factors like the network structure, the classification type, the used dataset. At the end of the survey, we will mention the merits and demerits of each approach and suggest some enhancements to avoid the noticed drawbacks.

The deep learning-based compressive Sensing (CS) has shown substantial improved performance and in run-time reduction with signal sampling and reconstruction. In most cases, moreover, these techniques suffer from disrupting artefacts or high-frequency contents at low sampling ratios. Similarly, this occurs in the multi-resolution sampling method, which further collects more components with lower frequencies. A promising innovation combining CS with convolutionary neural network has eliminated the sparsity constraint yet recovery persists slow. We propose a Deep wavelet based compressive sensing with multi-resolution framework provides better improvement in reconstruction as well as run time. The proposed model demonstrates outstanding quality on test functions over previous approaches.

This letter presents a novel median filtering forensics approach, based on a convolutional neural network (CNN) with an adaptive filtering layer (AFL), which is built in the discrete cosine transform (DCT) domain. Using the proposed AFL, the CNN can determine the main frequency range closely related with the operational traces. Then, to automatically learn the multi-scale manipulation features, a multi-scale convolutional block is developed, exploring a new multi-scale feature fusion strategy based on the maxout function. The resultant features are further processed by a convolutional stream with pooling and batch normalization operations, and finally fed into the classification layer with the Softmax function. Experimental results show that our proposed approach is able to accurately detect the median filtering manipulation and outperforms the state-of-the-art schemes, especially in the scenarios of low image resolution and serious compression loss.

In image forensics, to determine whether the image is impurely transformed, it extracts and examines the features included in the suspicious image. In general, the features extracted for the detection of forgery images are based on numerical values, so it is somewhat unreasonable to use in the CNN structure for image classification. In this paper, the extraction method of a feature vector is using a least-squares solution. Treat a suspicious image like a matrix and its solution to be coefficients as the feature vector. Get two solutions from two images of the original and its median filter residual (MFR). Subsequently, the two features were formed into a visualized pattern and then fed into CNN deep learning to classify the various transformed images. A new structure of the CNN net layer was also designed by hybrid with the inception module and the residual block to classify visualized feature vector patterns. The performance of the proposed image forensics detection (IFD) scheme was measured with the seven transformed types of image: average filtered (window size: 3 × 3), gaussian filtered (window size: 3 × 3), JPEG compressed (quality factor: 90, 70), median filtered (window size: 3 × 3, 5 × 5), and unaltered. The visualized patterns are fed into the image input layer of the designed CNN hybrid model. Throughout the experiment, the accuracy of median filtering detection was 98% over. Also, the area under the curve (AUC) by sensitivity (TP: true positive rate) and 1-specificity (FP: false positive rate) results of the proposed IFD scheme approached to `1' on the designed CNN hybrid model. Experimental results show high efficiency and performance to classify the various transformed images. Therefore, the grade evaluation of the proposed scheme is “Excellent (A)”.

Motions of facial components convey significant information of facial expressions. Although remarkable advancement has been made, the dynamic of facial topology has not been fully exploited. In this paper, a novel facial expression recognition (FER) algorithm called Spatial Temporal Semantic Graph Network (STSGN) is proposed to automatically learn spatial and temporal patterns through end-to-end feature learning from facial topology structure. The proposed algorithm not only has greater discriminative power to capture the dynamic patterns of facial expression and stronger generalization capability to handle different variations but also higher interpretability. Experimental evaluation on two popular datasets, CK+ and Oulu-CASIA, shows that our algorithm achieves more competitive results than other state-of-the-art methods.

The rapid growth of artificial intelligence has contributed a lot to the technology world. As the traditional algorithms failed to meet the human needs in real time, Machine learning and deep learning algorithms have gained great success in different applications such as classification systems, recommendation systems, pattern recognition etc. Emotion plays a vital role in determining the thoughts, behaviour and feeling of a human. An emotion recognition system can be built by utilizing the benefits of deep learning and different applications such as feedback analysis, face unlocking etc. can be implemented with good accuracy. The main focus of this work is to create a Deep Convolutional Neural Network (DCNN) model that classifies 5 different human facial emotions. The model is trained, tested and validated using the manually collected image dataset.

Due to the noise in the images, the edges extracted from these noisy images are always discontinuous and inaccurate by traditional operators. In order to solve these problems, this paper proposes multi-direction edge detection operator to detect edges from noisy images. The new operator is designed by introducing the shear transformation into the traditional operator. On the one hand, the shear transformation can provide a more favorable treatment for directions, which can make the new operator detect edges in different directions and overcome the directional limitation in the traditional operator. On the other hand, all the single pixel edge images in different directions can be fused. In this case, the edge information can complement each other. The experimental results indicate that the new operator is superior to the traditional ones in terms of the effectiveness of edge detection and the ability of noise rejection.

Malware refers to any software accessing or being installed in a system without the authorisation of administrators. Various malware has been widely used for cyber-criminals to accomplish their evil intentions and goals. To combat the increasing amount and reduce the threat of malicious programs, a novel deep learning framework, which uses NLP techniques for reference, combines CNN and LSTM neurones to capture the locally spatial correlations and learn from sequential longterm dependency is proposed. Hence, high-level abstractions and representations are automatically extracted for the malware classification task. The classification accuracy improves from 0.81 (best one by Random Forest) to approximately 1.0.

Traffic flow prediction is an important foundation for intelligent transportation systems. The traffic data are generated from a traffic network and evolved dynamically. So spatio-temporal relation exploration plays a support role on traffic data analysis. Most researches focus on spatio-temporal information fusion through a convolution operation. To the best of our knowledge, this is the first work to suggest that it is necessary to distinguish the two aspects of spatial correlations and propose the two types of spatial graphs, named as geographic graph and semantic graph. Then two novel stereo convolutions with irregular acceptive fields are proposed. The geographic-semantic-temporal contexts are dynamically jointly captured through performing the proposed convolutions on graph sequences. We propose a geographic-semantic-temporal graph convolutional network (GST-GCN) model that combines our graph convolutions and GRU units hierarchically in a unified end-to-end network. The experiment results on the Caltrans Performance Measurement System (PeMS) dataset show that our proposed model significantly outperforms other popular spatio-temporal deep learning models and suggest the effectiveness to explore geographic-semantic-temporal dependencies on deep learning models for traffic flow prediction.

Deepfake imagery that contains altered faces has become a threat to online content. Current anti-deepfake approaches usually do so by detecting image anomalies, such as visible artifacts or inconsistencies. However, with deepfake advances, these visual artifacts are becoming harder to detect. In this paper, we show that one can use biometric eyebrow matching as a tool to detect manipulated faces. Our method could provide an 0.88 AUC and 20.7% EER for deepfake detection when applied to the highest quality deepfake dataset, Celeb-DF.

Style transfer is a research hotspot in computer vision. Up to now, it is still a challenge although many researches have been conducted on it for high quality style transfer. In this work, we propose an algorithm named ASTCNN which is a real-time Arbitrary Style Transfer Convolution Neural Network. The ASTCNN consists of two independent encoders and a decoder. The encoders respectively extract style and content features from style and content and the decoder generates the style transferred image images. Experimental results show that ASTCNN achieves higher quality output image than the state-of-the-art style transfer algorithms and the floating point computation of ASTCNN is 23.3% less than theirs.

Over the years, technology has reformed the perception of the world related to security concerns. To tackle security problems, we proposed a system capable of detecting security alerts. System encompass audio events that occur as an outlier against background of unusual activity. This ambiguous behaviour can be handled by auditory classification. In this paper, we have discussed two techniques of extracting features from sound data including: time-based and signal based features. In first technique, we preserve time-series nature of sound, while in other signal characteristics are focused. Convolution neural network is applied for categorization of sound. Major aim of research is security challenges, so we have generated data related to surveillance in addition to available datasets such as UrbanSound 8k and ESC-50 datasets. We have achieved 94.6% accuracy for proposed methodology based on self-generated dataset. Improved accuracy on locally prepared dataset demonstrates novelty in research.

Attacks by Jamming on wireless communication network can provoke Denial of Services. According to the communication system which is affected, the consequences can be more or less critical. In this paper, we propose to develop an algorithm which could be implemented at the reception stage of a communication terminal in order to detect the presence of jamming signals. The work is performed on Wi-Fi communication signals and demonstrates the necessity to have a specific signal processing at the reception stage to be able to detect the presence of jamming signals.

Neural style transfer has drawn broad attention in recent years. However, most existing methods aim to explicitly model the transformation between different styles, and the learned model is thus not generalizable to new styles. We here attempt to separate the representations for styles and contents, and propose a generalized style transfer network consisting of style encoder, content encoder, mixer and decoder. The style encoder and content encoder are used to extract the style and content factors from the style reference images and content reference images, respectively. The mixer employs a bilinear model to integrate the above two factors and finally feeds it into a decoder to generate images with target style and content. To separate the style features and content features, we leverage the conditional dependence of styles and contents given an image. During training, the encoder network learns to extract styles and contents from two sets of reference images in limited size, one with shared style and the other with shared content. This learning framework allows simultaneous style transfer among multiple styles and can be deemed as a special 'multi-task' learning scenario. The encoders are expected to capture the underlying features for different styles and contents which is generalizable to new styles and contents. For validation, we applied the proposed algorithm to the Chinese Typeface transfer problem. Extensive experiment results on character generation have demonstrated the effectiveness and robustness of our method.

Humans have created many pioneers of art from the beginning of time. There are not many notable achievements by an artificial intelligence to create something visually captivating in the field of art. However, some breakthroughs were made in the past few years by learning the differences between the content and style of an image using convolution neural networks and texture synthesis. But most of the approaches have the limitations on either processing time, choosing a certain style image or altering the weight ratio of style image. Therefore, we are to address these restrictions and provide a system which allows any style image selection with a user defined style weight ratio in minimum time possible.

The face recognition methods based on convolutional neural network have achieved great success. The existing model usually used the residual network as the core architecture. The residual network is good at reusing features, but it is difficult to explore new features. And the densely connected network can be used to explore new features. We proposed a face recognition model named Dense Face to explore the performance of densely connected network in face recognition. The model is based on densely connected convolutional neural network and composed of Dense Block layers, transition layers and classification layer. The model was trained with the joint supervision of center loss and softmax loss through feature normalization and enabled the convolutional neural network to learn more discriminative features. The Dense Face model was trained using the public available CASIA-WebFace dataset and was tested on the LFW and the CAS-PEAL-Rl datasets. Experimental results showed that the densely connected convolutional neural network has achieved higher face verification accuracy and has better robustness than other model such as VGG Face and ResNet model.

It is challenging to detect weak hidden information in a JPEG compressed image. In this paper, we propose a 32-layer convolutional neural networks (CNNs) with feature reuse by concatenating all features from previous layers. The proposed method can improve the flow of gradient and information, and the shared features and bottleneck layers in the proposed CNN model further reduce the number of parameters dramatically. The experimental results shown that the proposed method significantly reduce the detection error rate compared with the existing JPEG steganalysis methods, e.g. state-of-the-art XuNet method and the conventional SCA-GFR method. Compared with XuNet method and conventional method SCA-GFR in detecting J-UNIWARD at 0.1 bpnzAC (bit per non-zero AC DCT coefficient), the proposed method can reduce detection error rate by 4.33% and 6.55% respectively.

With the continuous development of the Internet of Vehicles, vehicles are no longer isolated nodes, but become a node in the car network. The open Internet will introduce traditional security issues into the Internet of Things. In order to ensure the safety of the networked cars, we hope to set up an intrusion detection system (IDS) on the vehicle terminal to detect and intercept network attacks. In our work, we designed an intrusion detection system for the Internet of Vehicles based on a convolutional neural network, which can run in a low-powered embedded vehicle terminal to monitor the data in the car network in real time. Moreover, for the case of packet encryption in some car networks, we have also designed a separate version for intrusion detection by analyzing the packet header. Experiments have shown that our system can guarantee high accuracy detection at low latency for attack traffic.

As an emerging paradigm for energy-efficiency design, approximate computing can reduce power consumption through simplification of logic circuits. Although calculation errors are caused by approximate computing, their impacts on the final results can be negligible in some error resilient applications, such as Convolutional Neural Networks (CNNs). Therefore, approximate computing has been applied to CNNs to reduce the high demand for computing resources and energy. Compared with the traditional method such as reducing data precision, this paper investigates the effect of approximate computing on the accuracy and power consumption of CNNs. To optimize the approximate computing technology applied to CNNs, we propose a method for quantifying the error resilience of each neuron by theoretical analysis and observe that error resilience varies widely across different neurons. On the basic of quantitative error resilience, dynamic adaptation of approximate bit-width and the corresponding configurable adder are proposed to fully exploit the error resilience of CNNs. Experimental results show that the proposed method further improves the performance of power consumption while maintaining high accuracy. By adopting the optimal approximate bit-width for each layer found by our proposed algorithm, dynamic adaptation of approximate bit-width reduces power consumption by more than 30% and causes less than 1% loss of the accuracy for LeNet-5.

We investigate a deep learning model for action recognition that simultaneously extracts spatio-temporal information from a raw RGB input data. The proposed multiple spatio-temporal scales recurrent neural network (MSTRNN) model is derived by combining multiple timescale recurrent dynamics with a conventional convolutional neural network model. The architecture of the proposed model imposes both spatial and temporal constraints simultaneously on its neural activities. The constraints vary, with multiple scales in different layers. As suggested by the principle of upward and downward causation, it is assumed that the network can develop a functional hierarchy using its constraints during training. To evaluate and observe the characteristics of the proposed model, we use three human action datasets consisting of different primitive actions and different compositionality levels. The performance capabilities of the MSTRNN model on these datasets are compared with those of other representative deep learning models used in the field. The results show that the MSTRNN outperforms baseline models while using fewer parameters. The characteristics of the proposed model are observed by analyzing its internal representation properties. The analysis clarifies how the spatio-temporal constraints of the MSTRNN model aid in how it extracts critical spatio-temporal information relevant to its given tasks.

The answer selection task is one of the most important issues within the automatic question answering system, and it aims to automatically find accurate answers to questions. Traditional methods for this task use manually generated features based on tf-idf and n-gram models to represent texts, and then select the right answers according to the similarity between the representations of questions and the candidate answers. Nowadays, many question answering systems adopt deep neural networks such as convolutional neural network (CNN) to generate the text features automatically, and obtained better performance than traditional methods. CNN can extract consecutive n-gram features with fixed length by sliding fixed-length convolutional kernels over the whole word sequence. However, due to the complex semantic compositionality of the natural language, there are many phrases with variable lengths and be composed of non-consecutive words in natural language, such as these phrases whose constituents are separated by other words within the same sentences. But the traditional CNN is unable to extract the variable length n-gram features and non-consecutive n-gram features. In this paper, we propose a multi-scale deformable convolutional neural network to capture the non-consecutive n-gram features by adding offset to the convolutional kernel, and also propose to stack multiple deformable convolutional layers to mine multi-scale n-gram features by the means of generating longer n-gram in higher layer. Furthermore, we apply the proposed model into the task of answer selection. Experimental results on public dataset demonstrate the effectiveness of our proposed model in answer selection.

While recent studies have shed light on the expressivity, complexity and compositionality of convolutional networks, the real inductive bias of the family of functions reachable by gradient descent on natural data is still unknown. By exploiting symmetries in the preactivation space of convolutional layers, we present preliminary empirical evidence of regularities in the preimage of trained rectifier networks, in terms of arrangements of polytopes, and relate it to the nonlinear transformations applied by the network to its input.