arXiv subjects

Qiang Xu

Publications and source records attributed to Qiang Xu.

At least 199 records · Page 11Linked to original sources

arXiv2019

I Know What You See: Power Side-Channel Attack on Convolutional Neural Network Accelerators

Deep learning has become the de-facto computational paradigm for various kinds of perception problems, including many privacy-sensitive applications such as online medical image analysis. No doubt to say, the data privacy of these deep learning systems is a serious concern. Different from previous research focusing on exploiting privacy leakage from deep learning models, in this paper, we present the first attack on the implementation of deep learning models. To be specific, we perform the attack on an FPGA-based convolutional neural network accelerator and we manage to recover the input image from the collected power traces without knowing the detailed parameters in the neural network. For the MNIST dataset, our power side-channel attack is able to achieve up to 89% recognition accuracy.

cs.CV↗

arXiv2019

On Functional Test Generation for Deep Neural Network IPs

Machine learning systems based on deep neural networks (DNNs) produce state-of-the-art results in many applications. Considering the large amount of training data and know-how required to generate the network, it is more practical to use third-party DNN intellectual property (IP) cores for many designs. No doubt to say, it is essential for DNN IP vendors to provide test cases for functional validation without leaking their parameters to IP users. To satisfy this requirement, we propose to effectively generate test cases that activate parameters as many as possible and propagate their perturbations to outputs. Then the functionality of DNN IPs can be validated by only checking their outputs. However, it is difficult considering large numbers of parameters and highly non-linearity of DNNs. In this paper, we tackle this problem by judiciously selecting samples from the DNN training set and applying a gradient-based method to generate new test cases. Experimental results demonstrate the efficacy of our proposed solution.

cs.LG↗

arXiv2018

Structure-Aware 3D Hourglass Network for Hand Pose Estimation from Single Depth Image

In this paper, we propose a novel structure-aware 3D hourglass network for hand pose estimation from a single depth image, which achieves state-of-the-art results on MSRA and NYU datasets. Compared to existing works that perform image-to-coordination regression, our network takes 3D voxel as input and directly regresses 3D heatmap for each joint. To be specific, we use hourglass network as our backbone network and modify it into 3D form. We explicitly model tree-like finger bone into the network as well as in the loss function in an end-to-end manner, in order to take the skeleton constraints into consideration. Final estimation can then be easily obtained from voxel density map with simple post-processing. Experimental results show that the proposed structure-aware 3D hourglass network is able to achieve a mean joint error of 7.4 mm in MSRA and 8.9 mm in NYU datasets, respectively.

cs.CV↗

arXiv2018

Convergence rates on periodic homogenization of p-Laplace type equations

In this paper, we find some error estimates for periodic homogenization of p-Laplace type equations under the same structure assumption on homogenized equations. The main idea is that by adjusting the size of the difference quotient of the correctors to make the convergence rate visible. In order to reach our goal, the corresponding flux corrector with some properties are developed. Meanwhile, the shift-arguments is in fact applied down to $\varepsilon$ scale, which leads to a new weighted type inequality for smoothing operator with the weight satisfying Harnack's inequality in small scales. As a result, it is possible to develop some large-scale estimates. We finally mention that our approach brought in a systematic error (this phenomenon will disappear in linear and non-degenerated cases), which was fortunately a quantity $o(\varepsilon)$ here.

math.AP↗

arXiv2018

On Configurable Defense against Adversarial Example Attacks

Machine learning systems based on deep neural networks (DNNs) have gained mainstream adoption in many applications. Recently, however, DNNs are shown to be vulnerable to adversarial example attacks with slight perturbations on the inputs. Existing defense mechanisms against such attacks try to improve the overall robustness of the system, but they do not differentiate different targeted attacks even though the corresponding impacts may vary significantly. To tackle this problem, we propose a novel configurable defense mechanism in this work, wherein we are able to flexibly tune the robustness of the system against different targeted attacks to satisfy application requirements. This is achieved by refining the DNN loss function with an attack sensitive matrix to represent the impacts of different targeted attacks. Experimental results on CIFAR-10 and GTSRB data sets demonstrate the efficacy of the proposed solution.

cs.LG↗

arXiv2018

Invocation-driven Neural Approximate Computing with a Multiclass-Classifier and Multiple Approximators

Neural approximate computing gains enormous energy-efficiency at the cost of tolerable quality-loss. A neural approximator can map the input data to output while a classifier determines whether the input data are safe to approximate with quality guarantee. However, existing works cannot maximize the invocation of the approximator, resulting in limited speedup and energy saving. By exploring the mapping space of those target functions, in this paper, we observe a nonuniform distribution of the approximation error incurred by the same approximator. We thus propose a novel approximate computing architecture with a Multiclass-Classifier and Multiple Approximators (MCMA). These approximators have identical network topologies and thus can share the same hardware resource in a neural processing unit(NPU) clip. In the runtime, MCMA can swap in the invoked approximator by merely shipping the synapse weights from the on-chip memory to the buffers near MAC within a cycle. We also propose efficient co-training methods for such MCMA architecture. Experimental results show a more substantial invocation of MCMA as well as the gain of energy-efficiency.

cs.LG↗

arXiv2018

Quantitative Estimates on Periodic Homogenization of Nonlinear Elliptic Operators

In this paper, we are interested in the periodic homogenization of quasilinear elliptic equations. We obtain error estimates $O(\varepsilon^{1/2})$ for a $C^{1,1}$ domain, and $O(\varepsilon^σ)$ for a Lipschitz domain, in which $σ\in(0,1/2)$ is close to zero. Based upon the convergence rates, an interior Lipschitz estimate, as well as a boundary Hölder estimate can be developed at large scales without any smoothness assumption, and these will implies reverse Hölder estimates established for a $C^1$ domain. By a real method developed by Z.Shen \cite{S3}, we consequently derive a global $W^{1,p}$ estimate for $2\leq p<\infty$. This work may be regarded as an extension of \cite{MAFHL,S5} to a nonlinear operator, and our results may be extended to the related Neumann boundary problems without any real difficulty.

math.AP↗

arXiv2018

$L^p$ Neumann Problems in Homogenization of General Elliptic Systems

In this paper, we extend the nontangential maximal function estimate obtained by C. Kenig, F. Lin and Z. Shen in \cite{KFS1} to the nonhomogeneous elliptic operators with rapidly oscillating periodic coefficients. The result relies on the previous work \cite{X2}, and optimal boundary estimates which is based upon certain estimates on convergence rates. Compared to the homogeneous case, the additional bootstrap process seems inevitable, and the Neumann boundary corrector caused by the lower order term are still useful here.

math.AP↗

arXiv2018

The Methods of Layer Potentials for General Elliptic Homogenization Problems in Lipschitz Domains

In terms of layer potential methods, this paper is devoted to study the $L^2$ boundary value problems for nonhomogeneous elliptic operators with rapidly oscillating coefficients in a periodic setting. Under a low regularity assumption on the coefficients, we establish the solvability for Dirichlet, regular and Neumann problems in a bounded Lipschitz domain, as well as, the uniform nontangential maximal function estimates and square function estimates. The main difficulty is reflected in two aspects: (i) we can not treat the lower order terms as a compact perturbation to the leading term due to the low regularity assumption; (ii) the nonhomogeneous systems do not possess a scaling-invariant property in general. Although this work may be regarded as a follow-up to C. Kenig and Z. Shen's in \cite{SZW24}, we make an effort to find a clear way of how to handle the nonhomogeneous operators by using the known results of the homogenous ones. Also, we mention that the periodicity condition plays a key role in the scaling-invariant estimates.

math.AP↗

arXiv2018

Towards Imperceptible and Robust Adversarial Example Attacks against Neural Networks

Machine learning systems based on deep neural networks, being able to produce state-of-the-art results on various perception tasks, have gained mainstream adoption in many applications. However, they are shown to be vulnerable to adversarial example attack, which generates malicious output by adding slight perturbations to the input. Previous adversarial example crafting methods, however, use simple metrics to evaluate the distances between the original examples and the adversarial ones, which could be easily detected by human eyes. In addition, these attacks are often not robust due to the inevitable noises and deviation in the physical world. In this work, we present a new adversarial example attack crafting method, which takes the human perceptual system into consideration and maximizes the noise tolerance of the crafted adversarial example. Experimental results demonstrate the efficacy of the proposed technique.

cs.LG↗

arXiv2017

Detecting Location Fraud in Indoor Mobile Crowdsensing

Mobile crowdsensing allows a large number of mobile devices to measure phenomena of common interests and form a body of knowledge about natural and social environments. In order to get location annotations for indoor mobile crowdsensing, reference tags are usually deployed which are susceptible to tampering and compromises by attackers. In this work, we consider three types of location-related attacks including tag forgery, tag misplacement, and tag removal. Different detection algorithms are proposed to deal with these attacks. First, we introduce location-dependent fingerprints as supplementary information for better location identification. A truth discovery algorithm is then proposed to detect falsified data. Moreover, visiting patterns are utilized for the detection of tag misplacement and removal. Experiments on both crowdsensed and emulated dataset show that the proposed algorithms can detect all three types of attacks with high accuracy.

cs.CR↗

arXiv2017

TuRF: Fast Data Collection for Fingerprint-based Indoor Localization

Many infrastructure-free indoor positioning systems rely on fine-grained location-dependent fingerprints to train models for localization. The site survey process to collect fingerprints is laborious and is considered one of the major obstacles to deploying such systems. In this paper, we propose TuRF, a fast path-based fingerprint collection mechanism for site survey. We demonstrate the feasibility to collect fingerprints for indoor localization during walking along predefined paths. A step counter is utilized to accommodate the variations in walking speed. Approximate location labels inferred from the steps are then used to train a Gaussian Process regression model. Extensive experiments show that TuRF can significantly reduce the required time for site survey, without compromising the localization performance.

cs.NI↗

arXiv2017

Quantitative Estimates in Homogenization of Parabolic Systems of Elasticity in Lipschitz Cylinders

In a Lipschitz cylinder, this paper is devoted to establish an almost sharp error estimate $O(\varepsilon\log_2(1/\varepsilon))$ in $L^2$-norm for parabolic systems of elasticity with initial-Dirichlet conditions, arising in the homogenization theory. To achieve the goal, with the parabolic distance function being a weight, we first developed some new weighted-type inequalities for the smoothing operator at scale $\varepsilon$ in terms of t-anisotropic Sobolev spaces, and then all the problems may be reduced to three kinds of estimate for the homogenized system, in which a weighted-type Caccioppoli's inequality on time-layer has also been found. Throughout the paper, we do not introduce any smoothness on coefficients compared to the arguments investigated by C.Kenig, F. Lin and Z. Shen in \cite{SZW2}, while this study can be considered to be a further development of \cite{GZS} and \cite{QX2}.

math.AP↗

arXiv2017

Commutator Estimates for the Dirichlet-to-Neumann Map of Stokes Systems in Lipschitz Domains

In the paper, we establish commutator estimates for the Dirichlet-to-Neumann map of Stokes systems in Lipschitz domains. The approach is based on Dahlberg's bilinear estimates, and the results may be regarded as an extension of [Dahlberg, Poisson semigroups and singular integrals, Proc. Amer. Math. Soc., 97(1986), no.1, 41-48.] and [Shen, Commutator estimates for the Dirichlet-to-Neumann map in Lipschitz domains, Some topics in harmonic analysis and applications, 369C384, Adv. Lect. Math. (ALM)34, Int. Press, Somerville, MA, 2016.] to Stokes systems.

math.AP↗

arXiv2016

Optimal Boundary Estimates for Stokes Systems in Homogenization Theory

The paper concerns the sharp boundary regularity estimates in homogenization of Dirichlet problem for Stokes systems. We obtain the Lipschitz estimates for velocity term and $L^\infty$ estimate for pressure term, under some reasonable smoothness assumption on rapidly oscillating periodic coefficients. The approach is based on convergence rates, originally investigated by S. Armstrong and Z. Shen in \cite{SZ,SZW12}, however the argument developed here does not rely on the Rellich estimates. In this sense, we find a new way to obtain the sharp uniform boundary estimates without imposing the symmetry assumption on coefficients. Additionally, we emphasize that $L^\infty$ estimate for the pressure term does require the $O(\varepsilon^{1/2})$ convergence rate, locally at least, compared to $O(\varepsilon^λ)$ for the velocity term, where $λ\in(0,1/2)$.

math.AP↗

arXiv2016

Convergence rates for general elliptic homogenization problems in a bounded Lipschitz domain

The paper extends the results obtained by C. Kenig, F. Lin and Z. Shen in \cite{SZW2} to more general elliptic homogenization problems in two perspectives: lower order terms in the operator and no smoothness on the coefficients. We do not repeat their arguments. Instead we find the new weighted-type estimates for the smoothing operator at scale $\varepsilon$, and combining some techniques developed by Z. Shen in \cite{SZW12} leads to our main results. In addition, we also obtain sharp $O(\varepsilon)$ convergence rates in $L^{p}$ with $p=2d/(d-1)$, which were originally established by Z. Shen for elasticity systems in \cite{SZW12}. Also, this work may be regarded as the extension of \cite{TS,TS2} developed by T. Suslina concerned with the bounded Lipschitz domain.

math.AP↗

arXiv2016

Convergence rates and $W^{1,p}$ estimates in homogenization theory of Stokes systems in Lipschitz domains

Concerned with the Stokes systems with rapidly oscillating periodic coefficients, we mainly extend the recent works in \cite{SGZWS,G} to those in term of Lipschitz domains. The arguments employed here are quite different from theirs, and the basic idea comes from \cite{QX2}, originally motivated by \cite{SZW2,SZW12,TS}. We obtain an almost-sharp $O(\varepsilon\ln(r_0/\varepsilon))$ convergence rate in $L^2$ space, and a sharp $O(\varepsilon)$ error estimate in $L^{\frac{2d}{d-1}}$ space by a little stronger assumption. Under the dimensional condition $d=2$, we also establish the optimal $O(\varepsilon)$ convergence rate on pressure terms in $L^2$ space. Then utilizing the convergence rates we can derive the $W^{1,p}$ estimates uniformly down to microscopic scale $\varepsilon$ without any smoothness assumption on the coefficients, where $|\frac{1}{p}-\frac{1}{2}|<\frac{1}{2d}+ε$ and $ε$ is a positive constant independent of $\varepsilon$. Combining the local estimates, based upon $\text{VMO}$ coefficients, consequently leads to the uniform $W^{1,p}$ estimates. Here the proofs do not rely on the well known compactness methods.

math.AP↗

arXiv2016

A linear scaling method to evaluate the ion-electron potential of crystalline solids

We propose a simple linear scaling expression in reciprocal space for evaluating the ion--electron potential of crystalline solids. The expression replaces the long-range ion--electron potential with an equivalent localized charge distribution and corresponding boundary conditions on the unit cell. Given that no quadratic scaling structure factor is required---as used in traditional methods---the expression shows inherent linear behavior, and is well suited to simulating large-scale systems within orbital-free density functional theory. The scheme is implemented in the ATLAS software package and benchmarked by using a solid Mg bcc lattice containing tens of thousands of atoms in the unit cell. The test results show that the method can efficiently model large crystals with high computational accuracy.

cond-mat.mtrl-sci↗