Biblio

Open set recognition (OSR) problem has been a challenge in many machine learning (ML) applications, such as security. As new/unknown malware families occur regularly, it is difficult to exhaust samples that cover all the classes for the training process in ML systems. An advanced malware classification system should classify the known classes correctly while sensitive to the unknown class. In this paper, we introduce a self-supervised pre-training approach for the OSR problem in malware classification. We propose two transformations for the function call graph (FCG) based malware representations to facilitate the pretext task. Also, we present a statistical thresholding approach to find the optimal threshold for the unknown class. Moreover, the experiment results indicate that our proposed pre-training process can improve different performances of different downstream loss functions for the OSR problem.

With the development of network technology, identifying specific traffic has become important in network monitoring and security. However, designing feature sets that can accurately describe network traffic is still an urgent problem. Most of existing researches cannot realize effectively the identification of targets, and don't perform well in the complex and dynamic network environment. Aiming at these problems, we propose a novel method in this paper, which learns correlation features of network traffic based on the hierarchical structure. Firstly, the method learns the spatial-temporal features using convolutional neural networks (CNNs) and the bidirectional long short-term memory networks (Bi-LSTMs), then builds network topology to capture dependency characteristics between sessions and learns the context-related features through the graph attention networks (GATs). Finally, the network traffic session is classified using a fully connected network. The experimental results show that our method can effectively improve the detection ability and achieve a better classification performance overall.

In recent years, hashing algorithm has been widely researched and has made considerable progress in large-scale image retrieval tasks due to its advantages of convenient storage and fast calculation efficiency. Nowadays most researchers use deep convolutional neural networks (CNNs) to perform feature learning and hash coding learning at the same time for image retrieval and the deep hashing methods based on deep CNNs perform much better than the traditional manual feature hashing methods. But most methods are designed to handle simple binary similarity and decrease quantization error, ignoring that the features of similar images and hashing codes generated are not compact enough. In order to enhance the performance of CNNs-based hashing algorithms for large scale image retrieval, this paper proposes a new deep-supervised hashing algorithm in which a novel channel attention mechanism is added and the loss function is elaborately redesigned to generate compact binary codes. It experimentally proves that, compared with the existing hashing methods, this method has better performance on two large scale image datasets CIFAR-10 and NUS-WIDE.

Language creates a compact representation of the world and allows the description of unlimited situations and objectives through compositionality. While these characterizations may foster instructing, conditioning or structuring interactive agent behavior, it remains an open-problem to correctly relate language understanding and reinforcement learning in even simple instruction following scenarios. This joint learning problem is alleviated through expert demonstrations, auxiliary losses, or neural inductive biases. In this paper, we propose an orthogonal approach called Hindsight Generation for Experience Replay (HIGhER) that extends the Hindsight Experience Replay approach to the language-conditioned policy setting. Whenever the agent does not fulfill its instruction, HIGhER learns to output a new directive that matches the agent trajectory, and it relabels the episode with a positive reward. To do so, HIGhER learns to map a state into an instruction by using past successful trajectories, which removes the need to have external expert interventions to relabel episodes as in vanilla HER. We show the efficiency of our approach in the BabyAI environment, and demonstrate how it complements other instruction following methods.

Intrusion detection system based on representation learning is the main research direction in the field of anomaly detection. Malicious traffic detection system can distinguish normal and malicious traffic by learning representations between normal and malicious traffic. However, under the context of big data, there are many types of malicious traffic, and the features are also changing constantly. It is still a urgent problem to design a detection model that can effectively learn and summarize the feature of normal traffic and accurately identify the features of new kinds of malicious traffic.in this paper, a malicious traffic detection method based on Deep Support Vector Data Description is proposed, which is called Deep - SVDD. We combine convolutional neural network (CNN) with support vector data description, and train the model with normal traffic. The normal traffic features are mapped to high-dimensional space through neural networks, and a compact hypersphere is trained by unsupervised learning, which includes the normal features of the highdimensional space. Malicious traffic fall outside the hypersphere, thus distinguishing between normal and malicious traffic. Experiments show that the model has a high detection rate and a low false alarm rate, and it can effectively identify new malicious traffic.

Neural architectures are the foundation for improving performance of deep neural networks (DNNs). This paper presents deep compositional grammatical architectures which harness the best of two worlds: grammar models and DNNs. The proposed architectures integrate compositionality and reconfigurability of the former and the capability of learning rich features of the latter in a principled way. We utilize AND-OR Grammar (AOG) as network generator in this paper and call the resulting networks AOGNets. An AOGNet consists of a number of stages each of which is composed of a number of AOG building blocks. An AOG building block splits its input feature map into N groups along feature channels and then treat it as a sentence of N words. It then jointly realizes a phrase structure grammar and a dependency grammar in bottom-up parsing the “sentence” for better feature exploration and reuse. It provides a unified framework for the best practices developed in state-of-the-art DNNs. In experiments, AOGNet is tested in the ImageNet-1K classification benchmark and the MS-COCO object detection and segmentation benchmark. In ImageNet-1K, AOGNet obtains better performance than ResNet and most of its variants, ResNeXt and its attention based variants such as SENet, DenseNet and DualPathNet. AOGNet also obtains the best model interpretability score using network dissection. AOGNet further shows better potential in adversarial defense. In MS-COCO, AOGNet obtains better performance than the ResNet and ResNeXt backbones in Mask R-CNN.

A plethora of recent work has shown that convolutional networks are not robust to adversarial images: images that are created by perturbing a sample from the data distribution as to maximize the loss on the perturbed example. In this work, we hypothesize that adversarial perturbations move the image away from the image manifold in the sense that there exists no physical process that could have produced the adversarial image. This hypothesis suggests that a successful defense mechanism against adversarial images should aim to project the images back onto the image manifold. We study such defense mechanisms, which approximate the projection onto the unknown image manifold by a nearest-neighbor search against a web-scale image database containing tens of billions of images. Empirical evaluations of this defense strategy on ImageNet suggest that it very effective in attack settings in which the adversary does not have access to the image database. We also propose two novel attack methods to break nearest-neighbor defense settings and show conditions under which nearest-neighbor defense fails. We perform a series of ablation experiments, which suggest that there is a trade-off between robustness and accuracy between as we use features from deeper in the network, that a large index size (hundreds of millions) is crucial to get good performance, and that careful construction of database is crucial for robustness against nearest-neighbor attacks.

Deep learning techniques have been widely used in intrusion detection, but their application on convolutional neural networks (CNN) is still immature. The main challenge is how to represent the network traffic to improve performance of the CNN model. In this paper, we propose a network intrusion detection algorithm based on representation learning using Fast Fourier Transformation (FFT), which is first exploration that converts traffic to image by FFT to the best of our knowledge. Each traffic is converted to an image and then the intrusion detection problem is turned to image classification. The experiment results on NSL-KDD dataset show that the classification performence of the algorithm in the CNN model has obvious advantages compared with other algorithms.

In this paper we present a new approach, named DLGraph, for malware detection using deep learning and graph embedding. DLGraph employs two stacked denoising autoencoders (SDAs) for representation learning, taking into consideration computer programs' function-call graphs and Windows application programming interface (API) calls. Given a program, we first use a graph embedding technique that maps the program's function-call graph to a vector in a low-dimensional feature space. One SDA in our deep learning model is used to learn a latent representation of the embedded vector of the function-call graph. The other SDA in our model is used to learn a latent representation of the given program's Windows API calls. The two learned latent representations are then merged to form a combined feature vector. Finally, we use softmax regression to classify the combined feature vector for predicting whether the given program is malware or not. Experimental results based on different datasets demonstrate the effectiveness of the proposed approach and its superiority over a related method.

Hierarchical approaches for representation learning have the ability to encode relevant features at multiple scales or levels of abstraction. However, most hierarchical approaches exploit only the last level in the hierarchy, or provide a multiscale representation that holds a significant amount of redundancy. We argue that removing redundancy across the multiple levels of abstraction is important for an efficient representation of compositionality in object-based representations. With the perspective of feature learning as a data compression operation, we propose a new greedy inference algorithm for hierarchical sparse coding. Convolutional matching pursuit with a L0-norm constraint was used to encode the input signal into compact and non-redundant codes distributed across levels of the hierarchy. Simple and complex synthetic datasets of temporal signals were created to evaluate the encoding efficiency and compare with the theoretical lower bounds on the information rate for those signals. Empirical evidence have shown that the algorithm is able to infer near-optimal codes for simple signals. However, it failed for complex signals with strong overlapping between objects. We explain the inefficiency of convolutional matching pursuit that occurred in such case. This brings new insights about the NP-hard optimization problem related to using L0-norm constraint in inferring optimally compact and distributed object-based representations.

Autoencoders have been successful in learning meaningful representations from image datasets. However, their performance on text datasets has not been widely studied. Traditional autoencoders tend to learn possibly trivial representations of text documents due to their confoundin properties such as high-dimensionality, sparsity and power-law word distributions. In this paper, we propose a novel k-competitive autoencoder, called KATE, for text documents. Due to the competition between the neurons in the hidden layer, each neuron becomes specialized in recognizing specific data patterns, and overall the model can learn meaningful representations of textual data. A comprehensive set of experiments show that KATE can learn better representations than traditional autoencoders including denoising, contractive, variational, and k-sparse autoencoders. Our model also outperforms deep generative models, probabilistic topic models, and even word representation models (e.g., Word2Vec) in terms of several downstream tasks such as document classification, regression, and retrieval.