A Systematic View of Leakage Risks in Deep Neural Network Systems

As deep neural networks (DNNs) continue their reach into a wide range of application domains, the neural network architecture of DNN models becomes an increasingly sensitive subject, due to either intellectual property protection or risks of adversarial attacks. In observing the large gap between the architectural surfaces exploration and the model integrity study, this paper first presents the formulated schema of the model leakage risks. Then, we propose DeepSniffer, a learning-based model extraction framework, to obtain the complete model architecture information without any prior knowledge of the victim model. It is robust to architectural and system noises introduced by the complex memory hierarchy and diverse run-time system optimizations. Taking GPU platforms as a showcase, DeepSniffer performs model extraction by learning both the architecture-level execution features of kernels and the inter-layer temporal association information introduced by the common practice of DNN design. We demonstrate that DeepSniffer works experimentally in the context of an off-the-shelf Nvidia GPU platform running a variety of DNN models. The extracted models are directly helpful to the attempting of crafting adversarial inputs. The DeepSniffer project has been released in https://github.com/xinghu7788/DeepSniffer.

Download Full-text

A Systematic View of Leakage Risks in Deep Neural Network Systems

10.36227/techrxiv.14167493 ◽

2021 ◽

Author(s):

Xing Hu ◽

Ling Liang ◽

chen xiaobing ◽

Lei Deng ◽

Yu Ji ◽

...

Keyword(s):

Neural Network ◽

Network Architecture ◽

Temporal Association ◽

Neural Network Architecture ◽

Network Systems ◽

Model Extraction ◽

The Neural Network ◽

Wide Range ◽

The Common ◽

Nvidia Gpu

Download Full-text

Constrained Hybrid Neural Modelling of Biotechnological Processes

International Journal of Chemical Reactor Engineering ◽

10.2202/1542-6580.2117 ◽

2010 ◽

Vol 8 (1) ◽

Cited By ~ 2

Author(s):

Asma Karama ◽

Olivier Bernard ◽

Jean-Luc Gouzé

Keyword(s):

Neural Network ◽

Mass Balance ◽

Network Architecture ◽

A Priori ◽

Neural Model ◽

Neural Network Architecture ◽

The Neural Network ◽

Wide Range ◽

Neural Modelling ◽

Biotechnological Processes

We propose a general methodology to develop a hybrid neural model for a wide range of biotechnological processes. The hybrid neural modelling approach combines the flexibility of a neural network representation of unknown process kinetics with a global mass-balance based process description. The hybrid model is built in such a way that its trajectories keep their physical and biological meaning (mass balance, positivity of the concentrations, boundness, saturation or inhibition of kinetics) even far from the identification data conditions. We examine the constraints (a priori knowledge) that must be satisfied by the model and that provide additional conditions to be imposed on the neural network. We illustrate our approach with various biotechnological processes showing how to select the appropriate neural network architecture. The method is detailed for modelling an anaerobic wastewater treatment bioreactor using experimental data.

Download Full-text

SCORING MODELING BASED ON NEURAL NETWORKS FOR DETERMINING A BANK BORROWER'S RATING

Economy of Ukraine ◽

10.15407/economyukr.2020.10.054 ◽

2020 ◽

Vol 2020 (10) ◽

pp. 54-62

Author(s):

Oleksii VASYLIEV ◽

Keyword(s):

Neural Network ◽

Neural Networks ◽

Network Architecture ◽

Statistical Data ◽

Activation Function ◽

Decision Making Process ◽

Neural Network Architecture ◽

Acceptable Accuracy ◽

The Neural Network ◽

Sigmoid Activation Function

The problem of applying neural networks to calculate ratings used in banking in the decision-making process on granting or not granting loans to borrowers is considered. The task is to determine the rating function of the borrower based on a set of statistical data on the effectiveness of loans provided by the bank. When constructing a regression model to calculate the rating function, it is necessary to know its general form. If so, the task is to calculate the parameters that are included in the expression for the rating function. In contrast to this approach, in the case of using neural networks, there is no need to specify the general form for the rating function. Instead, certain neural network architecture is chosen and parameters are calculated for it on the basis of statistical data. Importantly, the same neural network architecture can be used to process different sets of statistical data. The disadvantages of using neural networks include the need to calculate a large number of parameters. There is also no universal algorithm that would determine the optimal neural network architecture. As an example of the use of neural networks to determine the borrower's rating, a model system is considered, in which the borrower's rating is determined by a known non-analytical rating function. A neural network with two inner layers, which contain, respectively, three and two neurons and have a sigmoid activation function, is used for modeling. It is shown that the use of the neural network allows restoring the borrower's rating function with quite acceptable accuracy.

Download Full-text

Spatial Variability Aware Deep Neural Networks (SVANN): A General Approach

ACM Transactions on Intelligent Systems and Technology ◽

10.1145/3466688 ◽

2021 ◽

Vol 12 (6) ◽

pp. 1-21

Author(s):

Jayant Gupta ◽

Carl Molnar ◽

Yiqun Xie ◽

Joe Knight ◽

Shashi Shekhar

Keyword(s):

Neural Network ◽

Spatial Variability ◽

Network Architecture ◽

Network Models ◽

Neural Network Architecture ◽

Neural Network Models ◽

Climatic Zones ◽

The Neural Network ◽

Plant Hardiness ◽

Interpretation Model

Spatial variability is a prominent feature of various geographic phenomena such as climatic zones, USDA plant hardiness zones, and terrestrial habitat types (e.g., forest, grasslands, wetlands, and deserts). However, current deep learning methods follow a spatial-one-size-fits-all (OSFA) approach to train single deep neural network models that do not account for spatial variability. Quantification of spatial variability can be challenging due to the influence of many geophysical factors. In preliminary work, we proposed a spatial variability aware neural network (SVANN-I, formerly called SVANN ) approach where weights are a function of location but the neural network architecture is location independent. In this work, we explore a more flexible SVANN-E approach where neural network architecture varies across geographic locations. In addition, we provide a taxonomy of SVANN types and a physics inspired interpretation model. Experiments with aerial imagery based wetland mapping show that SVANN-I outperforms OSFA and SVANN-E performs the best of all.

Download Full-text

Towards Better Interpretability in Deep Q-Networks

Proceedings of the AAAI Conference on Artificial Intelligence ◽

10.1609/aaai.v33i01.33014561 ◽

2019 ◽

Vol 33 ◽

pp. 4561-4569 ◽

Cited By ~ 2

Author(s):

Raghuram Mandyam Annasamy ◽

Katia Sycara

Keyword(s):

Neural Network ◽

Network Architecture ◽

Empirical Studies ◽

Superior Performance ◽

Training Algorithms ◽

Neural Network Architecture ◽

Q Learning ◽

The Neural Network ◽

Learning Techniques ◽

Out Of Sample

Deep reinforcement learning techniques have demonstrated superior performance in a wide variety of environments. As improvements in training algorithms continue at a brisk pace, theoretical or empirical studies on understanding what these networks seem to learn, are far behind. In this paper we propose an interpretable neural network architecture for Q-learning which provides a global explanation of the model’s behavior using key-value memories, attention and reconstructible embeddings. With a directed exploration strategy, our model can reach training rewards comparable to the state-of-the-art deep Q-learning models. However, results suggest that the features extracted by the neural network are extremely shallow and subsequent testing using out-of-sample examples shows that the agent can easily overfit to trajectories seen during training.

Download Full-text

Interpretable deep neural network for cancer survival analysis by integrating genomic and clinical data

BMC Medical Genomics ◽

10.1186/s12920-019-0624-2 ◽

2019 ◽

Vol 12 (S10) ◽

Cited By ~ 1

Author(s):

Jie Hao ◽

Youngsoon Kim ◽

Tejaswini Mallavarapu ◽

Jung Hun Oh ◽

Mingon Kang

Keyword(s):

Neural Network ◽

Survival Analysis ◽

Cancer Patient ◽

Clinical Data ◽

Network Architecture ◽

Deep Neural Network ◽

Patient Survival ◽

Biological Mechanisms ◽

Neural Network Architecture ◽

The Neural Network

Abstract Background Understanding the complex biological mechanisms of cancer patient survival using genomic and clinical data is vital, not only to develop new treatments for patients, but also to improve survival prediction. However, highly nonlinear and high-dimension, low-sample size (HDLSS) data cause computational challenges to applying conventional survival analysis. Results We propose a novel biologically interpretable pathway-based sparse deep neural network, named Cox-PASNet, which integrates high-dimensional gene expression data and clinical data on a simple neural network architecture for survival analysis. Cox-PASNet is biologically interpretable where nodes in the neural network correspond to biological genes and pathways, while capturing the nonlinear and hierarchical effects of biological pathways associated with cancer patient survival. We also propose a heuristic optimization solution to train Cox-PASNet with HDLSS data. Cox-PASNet was intensively evaluated by comparing the predictive performance of current state-of-the-art methods on glioblastoma multiforme (GBM) and ovarian serous cystadenocarcinoma (OV) cancer. In the experiments, Cox-PASNet showed out-performance, compared to the benchmarking methods. Moreover, the neural network architecture of Cox-PASNet was biologically interpreted, and several significant prognostic factors of genes and biological pathways were identified. Conclusions Cox-PASNet models biological mechanisms in the neural network by incorporating biological pathway databases and sparse coding. The neural network of Cox-PASNet can identify nonlinear and hierarchical associations of genomic and clinical data to cancer patient survival. The open-source code of Cox-PASNet in PyTorch implemented for training, evaluation, and model interpretation is available at: https://github.com/DataX-JieHao/Cox-PASNet.

Download Full-text

Neural Network Based Intelligent Learning of Fuzzy Logic Controller Parameters

Dynamic Systems and Control, Parts A and B ◽

10.1115/imece2004-59589 ◽

2004 ◽

Cited By ~ 8

Author(s):

Manish Kumar ◽

Devendra P. Garg

Keyword(s):

Neural Network ◽

Fuzzy Logic ◽

Network Architecture ◽

Fuzzy Logic Controller ◽

Rule Base ◽

Neural Network Architecture ◽

The Neural Network ◽

Neuro Fuzzy ◽

Learning Capabilities ◽

Optimal Signal

Design of an efficient fuzzy logic controller involves the optimization of parameters of fuzzy sets and proper choice of rule base. There are several techniques reported in recent literature that use neural network architecture and genetic algorithms to learn and optimize a fuzzy logic controller. This paper presents methodologies to learn and optimize fuzzy logic controller parameters that use learning capabilities of neural network. Concepts of model predictive control (MPC) have been used to obtain optimal signal to train the neural network via backpropagation. The strategies developed have been applied to control an inverted pendulum and results have been compared for two different fuzzy logic controllers developed with the help of neural networks. The first neural network emulates a PD controller, while the second controller is developed based on MPC. The proposed approach can be applied to learn fuzzy logic controller parameter online via the use of dynamic backpropagation. The results show that the Neuro-Fuzzy approaches were able to learn rule base and identify membership function parameters accurately.

Download Full-text

ORTHOSEG: A DEEP MULTIMODAL CONVOLUTONAL NEURAL NETWORK ARCHITECTURE FOR SEMANTIC SEGMENTATION OF ORTHOIMAGERY

ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences ◽

10.5194/isprs-archives-xlii-5-621-2018 ◽

2018 ◽

Vol XLII-5 ◽

pp. 621-628 ◽

Cited By ~ 2

Author(s):

P. Bodani ◽

K. Shreshtha ◽

S. Sharma

Keyword(s):

Neural Network ◽

Network Architecture ◽

Multiple Scales ◽

Semantic Segmentation ◽

Effective Field ◽

Surface Model ◽

Training Procedure ◽

Feature Maps ◽

Neural Network Architecture ◽

Wide Range

Abstract. This paper addresses the task of semantic segmentation of orthoimagery using multimodal data e.g. optical RGB, infrared and digital surface model. We propose a deep convolutional neural network architecture termed OrthoSeg for semantic segmentation using multimodal, orthorectified and coregistered data. We also propose a training procedure for supervised training of OrthoSeg. The training procedure complements the inherent architectural characteristics of OrthoSeg for preventing complex co-adaptations of learned features, which may arise due to probable high dimensionality and spatial correlation in multimodal and/or multispectral coregistered data. OrthoSeg consists of parallel encoding networks for independent encoding of multimodal feature maps and a decoder designed for efficiently fusing independently encoded multimodal feature maps. A softmax layer at the end of the network uses the features generated by the decoder for pixel-wise classification. The decoder fuses feature maps from the parallel encoders locally as well as contextually at multiple scales to generate per-pixel feature maps for final pixel-wise classification resulting in segmented output. We experimentally show the merits of OrthoSeg by demonstrating state-of-the-art accuracy on the ISPRS Potsdam 2D Semantic Segmentation dataset. Adaptability is one of the key motivations behind OrthoSeg so that it serves as a useful architectural option for a wide range of problems involving the task of semantic segmentation of coregistered multimodal and/or multispectral imagery. Hence, OrthoSeg is designed to enable independent scaling of parallel encoder networks and decoder network to better match application requirements, such as the number of input channels, the effective field-of-view, and model capacity.

Download Full-text

FID-Net: A Versatile Deep Neural Network Architecture for NMR Spectral Reconstruction and Virtual Decoupling

10.26434/chemrxiv.13295888 ◽

2020 ◽

Author(s):

Gogulan Karunanithy ◽

Flemming Hansen

Keyword(s):

Neural Network ◽

Network Architecture ◽

Nmr Spectra ◽

Deep Neural Network ◽

Protein Nmr ◽

Single Shot ◽

Neural Network Architecture ◽

Diverse Range ◽

Nmr Data ◽

Wide Range

In recent years, the transformative potential of deep neural networks (DNNs) for analysing and interpreting NMR data has clearly been recognised. However, most applications of DNNs in NMR to date either struggle to outperform existing methodologies or are limited in scope to a narrow range of data that closely resemble the data that the network was trained on. These limitations have prevented a widescale uptake of DNNs in NMR. Addressing this, we introduce FID-Net, a deep neural network architecture inspired by WaveNet, for performing analyses on time domain NMR data. We first demonstrate the effectiveness of this architecture in reconstructing non-uniformly sampled (NUS) biomolecular NMR spectra. It is shown that a single network is able to reconstruct a diverse range of 2D NUS spectra that have been obtained with arbitrary sampling schedules, with a range of sweep widths, and a variety of other acquisition parameters. The performance of the trained FID-Net in this case exceeds or matches existing methods currently used for the reconstruction of NUS NMR spectra. Secondly, we present a network based on the FID-Net architecture that can efficiently virtually decouple 13Cα-13Cβ couplings in HNCA protein NMR spectra in a single shot analysis, while at the same time leaving glycine residues unmodulated. The ability for these DNNs to work effectively in a wide range of scenarios, without retraining, paves the way for their widespread usage in analysing NMR data.

Download Full-text

Bioacoustic classification of avian calls from raw sound waveforms with an open-source deep learning architecture

Scientific Reports ◽

10.1038/s41598-021-95076-6 ◽

2021 ◽

Vol 11 (1) ◽

Author(s):

Francisco J. Bravo Sanchez ◽

Md Rahat Hossain ◽

Nathan B. English ◽

Steven T. Moore

Keyword(s):

Neural Network ◽

Network Architecture ◽

Limited Data ◽

Extraction Techniques ◽

Neural Network Architecture ◽

The Neural Network ◽

Current Classification ◽

Efficient Learning ◽

Sound Features

AbstractThe use of autonomous recordings of animal sounds to detect species is a popular conservation tool, constantly improving in fidelity as audio hardware and software evolves. Current classification algorithms utilise sound features extracted from the recording rather than the sound itself, with varying degrees of success. Neural networks that learn directly from the raw sound waveforms have been implemented in human speech recognition but the requirements of detailed labelled data have limited their use in bioacoustics. Here we test SincNet, an efficient neural network architecture that learns from the raw waveform using sinc-based filters. Results using an off-the-shelf implementation of SincNet on a publicly available bird sound dataset (NIPS4Bplus) show that the neural network rapidly converged reaching accuracies of over 65% with limited data. Their performance is comparable with traditional methods after hyperparameter tuning but they are more efficient. Learning directly from the raw waveform allows the algorithm to select automatically those elements of the sound that are best suited for the task, bypassing the onerous task of selecting feature extraction techniques and reducing possible biases. We use publicly released code and datasets to encourage others to replicate our results and to apply SincNet to their own datasets; and we review possible enhancements in the hope that algorithms that learn from the raw waveform will become useful bioacoustic tools.

Download Full-text