Biblio

Universal style transfer methods typically leverage rich representations from deep Convolutional Neural Network (CNN) models (e.g., VGG-19) pre-trained on large collections of images. Despite the effectiveness, its application is heavily constrained by the large model size to handle ultra-resolution images given limited memory. In this work, we present a new knowledge distillation method (named Collaborative Distillation) for encoder-decoder based neural style transfer to reduce the convolutional filters. The main idea is underpinned by a finding that the encoder-decoder pairs construct an exclusive collaborative relationship, which is regarded as a new kind of knowledge for style transfer models. Moreover, to overcome the feature size mismatch when applying collaborative distillation, a linear embedding loss is introduced to drive the student network to learn a linear embedding of the teacher's features. Extensive experiments show the effectiveness of our method when applied to different universal style transfer approaches (WCT and AdaIN), even if the model size is reduced by 15.5 times. Especially, on WCT with the compressed models, we achieve ultra-resolution (over 40 megapixels) universal style transfer on a 12GB GPU for the first time. Further experiments on optimization-based stylization scheme show the generality of our algorithm on different stylization paradigms. Our code and trained models are available at https://github.com/mingsun-tse/collaborative-distillation.

In this paper, we focus on developing a novel mechanism to preserve differential privacy in deep neural networks, such that: (1) The privacy budget consumption is totally independent of the number of training steps; (2) It has the ability to adaptively inject noise into features based on the contribution of each to the output; and (3) It could be applied in a variety of different deep neural networks. To achieve this, we figure out a way to perturb affine transformations of neurons, and loss functions used in deep neural networks. In addition, our mechanism intentionally adds "more noise" into features which are "less relevant" to the model output, and vice-versa. Our theoretical analysis further derives the sensitivities and error bounds of our mechanism. Rigorous experiments conducted on MNIST and CIFAR-10 datasets show that our mechanism is highly effective and outperforms existing solutions.

Security evaluation of diverse SDN frameworks is of significant importance to design resilient systems and deal with attacks. Focused on SDN scenarios, a game-theoretic model is proposed to analyze their security performance in existing SDN architectures. The model can describe specific traits in different structures, represent several types of information of players (attacker and defender) and quantitatively calculate systems' reliability. Simulation results illustrate dynamic SDN structures have distinct security improvement over static ones. Besides, effective dynamic scheduling mechanisms adopted in dynamic systems can enhance their security further.

In cyberspace, unknown zero-day attacks can bring safety hazards. Traditional defense methods based on signatures are ineffective. Based on the Cyberspace Mimic Defense (CMD) architecture, the paper proposes a framework to detect the attacks and respond to them. Inputs are assigned to all online redundant heterogeneous functionally equivalent modules. Their independent outputs are compared and the outputs in the majority will be the final response. The abnormal outputs can be detected and so can the attack. The damaged executive modules with abnormal outputs will be replaced with new ones from the diverse executive module pool. By analyzing the abnormal outputs, the correspondence between inputs and abnormal outputs can be built and inputs leading to recurrent abnormal outputs will be written into the zero-day attack related database and their reuses cannot work any longer, as the suspicious malicious inputs can be detected and processed. Further responses include IP blacklisting and patching, etc. The framework also uses honeypot like executive module to confuse the attacker. The proposed method can prevent the recurrent attack based on the same exploit.

The SDN (Software Defined Networking) paradigm rings flexibility to the network management and is an enabler to offer huge opportunities for network programmability. And, to solve the scalability issue raised by the centralized architecture of SDN, multi-controllers deployment (or distributed controllers system) is envisioned. In this paper, we focus on increasing the diversity of SDN control plane so as to enhance the network security. Our goal is to limit the ability of a malicious controller to compromise its neighboring controllers, and by extension, the rest of the controllers. We investigate a heterogeneous Susceptible-Infectious-Susceptible (SIS) epidemic model to evaluate the security performance and propose a coloring algorithm to increase the diversity based on community detection. And the simulation results demonstrate that our algorithm can reduce infection rate in control plane and our work shows that diversity must be introduced in network design for network security.