Summary
GANs with Tensor Networks Generative Adversarial Networks (GANs) have achieved impressive results in generating realistic images. However, existing GANs usually require a large amount of training data and high computational resources. Tensor factorization can effectively capture the latent structures in GANs and reduce the model size. By applying tensor factorization to the feature maps and parameters in GANs, the model size can be significantly reduced with little loss in generation quality. A novel tensor completion algorithm can be used as the generator in GANs to produce new training samples. The generated samples enrich the diversity of the training set and improve the detection efficiency and robustness of GANs. The generator in GANs can be adapted to multiple target domains using a hyper-network, which exploits the shared knowledge across domains and avoids linearly increasing computational requirements. The hyper-network dynamically adapts a pretrained GAN model to different target domains. The behavior of malware threats is increasing the need for malware detection. However, existing methods only target known samples, and detection of new malware and variants is limited. A GANs-based model can generate ‘true’ malware sample distributions to improve detection. Extracting texture features from Android malware APK files, and using tensor singular value decomposition to transform features into a third-order tensor for neural network training. A GANs model with the code tensor can surpass traditional models, improving efficiency up to 41.6%. Depth estimation from single images can be achieved by using an edge extraction network and dark channel prior in GANs. The edge network selects valid depth edges, and the dark channel generates a transmission map representing distance. The transmission map and RGB image are concatenated, and initial depth is generated through a generator network. Comparing the edge map of initial depth and input RGB selects valid edges to improve the generator. Experiments show cleared edges and better results than other methods. Location awareness for mobile computing can use RF fingerprint-based localization, though it is challenging. Modeling RF fingerprints as a 3D low-tubal-rank tensor captures multidimensional latent structures. A Tensor-GAN, with a generator using tensor completion to produce new fingerprints, and a regressor estimating locations, improves accuracy to 0.19m from 0.42m. Implemented on Android, the memory footprint is 57KB. Frequency regularization can be applied to CNN parameters by maintaining tensors in the frequency domain, where high frequencies are set to zero. Inverse DCT reconstructs the spatial tensors for network training. Since high frequency image components are less critical, many parameters can be set to zero. On various networks, over 1100x reduction is achieved for a 2% accuracy drop. A 34M parameter UNet was reduced to 4.5KB.
Recent advances in GANs have enabled visual editing and synthesis. However, existing methods either discover global semantics without localized control or require supervision.We present an unsupervised approach that discovers spatial parts and appearances. By applying tensor factorization to feature maps, we enable context-aware local editing with pixel control. The appearance factors correspond to saliency maps localizing concepts without labels. Experiments on GANs and datasets show our method is more efficient, providing accurate localized control.
Published By:
James Oldfield - International Conference on Learning Representations
2022
Cited By:
9
The proposed tensor virtualization technique efficiently reorganizes data with minimal hardware for CNN accelerators.
Published By:
Donghyun Kang - Design Automation Conference
2020
Cited By:
0
Recent work achieves photo-realistic generation using generative adversarial networks (GANs). Our method achieves better disentanglement and discovers semantic latent codes without supervision.
Published By:
Grigorios G. Chrysos - IEEE International Joint Conference on Neural Network
2020
Cited By:
1
GPU and TPU execution pattern analyzed. Bottleneck identified for GAN structure.
Published By:
A. Ravikumar - undefined
2022
Cited By:
0
We propose a depth estimation method combining edge extraction GAN and dark channel prior. Experiments show enhanced visual quality and metrics.
Published By:
Ying Li - IEEE Access
2021
Cited By:
4
Malware threats are increasing,requiring efficient detection.A model detects Android malware via tSVD, GANs and neural networks, improving detection by 41.6%.
Published By:
Zhao Yang - International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery
2022
Cited By:
0
Convolutional neural networks have shown impressive results.The proposed frequency regularization allows very significant parameter reduction with little accuracy loss.
Published By:
Chenqiu Zhao - arXiv.org
2023
Cited By:
0
Tensor virtualization technique allows adder-tree CNN accelerators to accelerate neural networks requiring data reorganization, including U-Net, DCGAN, and SRGAN.
Published By:
Donghyun Kang - Design Automation Conference
2020
Cited By:
0
The behavior of malware threats is gradually increasing, heightened the need for malware detection.We propose a novel scheme that detects malware and its variants efficiently. Based on generative adversarial networks (GANs), we obtain the ‘true’ sample distribution that satisfies the characteristics of real malware, use them to deceive the discriminator, thus achieve the defense against malicious code attacks and improve malware detection. A new Android malware APK to image texture feature extraction segmentation method is proposed, called segment self-growing texture segmentation algorithm. Tensor singular value decomposition (tSVD) based on the low-tubal rank transforms malicious features with different sizes into a fixed third-order tensor uniformly, which is entered into the neural network for training and learning. Finally, a flexible Android malware detection model based on GANs with code tensor (MTFD-GANs) is proposed. Experiments show that the proposed model can generally surpass the traditional malware detection model, with a maximum improvement efficiency of 41.6%. The newly generated samples of the GANs generator greatly enrich the sample diversity. And retraining malware detector can effectively improve the detection efficiency and robustness of traditional models.
Published By:
Zhao Yang - International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery
2022
Cited By:
0
We propose single image depth estimation using edge network and dark channel prior.
Published By:
Ying Li - IEEE Access
2021
Cited By:
4