Search

Deep neural networks (DNNs) have recently been widely adopted in various applications, and such success is largely due to a combination of algorithmic breakthroughs, computation resource improvements, and access to a large amount of data. However, the large-scale data collections required for deep learning often contain sensitive information, therefore raising many privacy concerns. Prior research has shown several successful attacks in inferring sensitive training data information, such as model inversion, membership inference, and generative adversarial networks (GAN) based leakage attacks against collaborative deep learning. In this paper, to enable learning efficiency as well as to generate data with privacy guarantees and high utility, we propose a differentially private autoencoder-based generative model (DP-AuGM) and a differentially private variational autoencoder-based generative model (DP-VaeGM). We evaluate the robustness of two proposed models. We show that DP-AuGM can effectively defend against the model inversion, membership inference, and GAN-based attacks. We also show that DP-VaeGM is robust against the membership inference attack. We conjecture that the key to defend against the model inversion and GAN-based attacks is not due to differential privacy but the perturbation of training data. Finally, we demonstrate that both DP-AuGM and DP-VaeGM can be easily integrated with real-world machine learning applications, such as machine learning as a service and federated learning, which are otherwise threatened by the membership inference attack and the GAN-based attack, respectively.

We report the first detection of >100 MeV gamma rays associated with a behind-the-limb solar flare, which presents a unique opportunity to probe the underlying physics of high-energy flare emission and particle acceleration. On 2013 October 11 a GOES M1.5 class solar flare occurred ~ 9.9 degrees behind the solar limb as observed by STEREO-B. RHESSI observed hard X-ray emission above the limb, most likely from the flare loop-top, as the footpoints were occulted. Surprisingly, the Fermi Large Area Telescope (LAT) detected >100 MeV gamma-rays for ~30 minutes with energies up to GeV. The LAT emission centroid is consistent with the RHESSI hard X-ray source, but its uncertainty does not constrain the source to be located there. The gamma-ray spectra can be adequately described by bremsstrahlung radiation from relativistic electrons having a relatively hard power-law spectrum with a high-energy exponential cutoff, or by the decay of pions produced by accelerated protons and ions with an isotropic pitch-angle distribution and a power-law spectrum with a number index of ~3.8. We show that high optical depths rule out the gamma rays originating from the flare site and a high-corona trap model requires very unusual conditions, so a scenario in which some of the particles accelerated by the CME shock travel to the visible side of the Sun to produce the observed gamma rays may be at work., Comment: 6 pages, 3 figures, accepted for publication on ApJ Letters

We have developed an inversion method for determination of the characteristics of the acceleration mechanism directly and non-parametrically from observations, in contrast to the usual forward fitting of parametric model variables to observations. In two recent papers (Petrosian & Chen 2010, Chen & Petrosian 2013), we demonstrate the efficacy of this inversion method by its application to acceleration of electrons in solar flares based on stochastic acceleration by turbulence. Here we explore its application for determining the characteristics of shock acceleration in supernova remnants (SNRs) based on the electron spectra deduced from the observed nonthermal radiation from SNRs and the spectrum of the cosmic ray electrons observed near the Earth. These spectra are related by the process of escape of the electrons from SNRs and energy loss during their transport in the galaxy. Thus, these observations allow us to determine spectral characteristics of the momentum and pitch angle diffusion coefficients, which play crucial roles in both direct acceleration by turbulence and in high Mach number shocks. Assuming that the average electron spectrum deduced from a few well known SNRs is representative of those in the solar neighborhood we find interesting discrepancies between our deduced forms for these coefficients and those expected from well known wave-particle interactions. This may indicate that the standard assumptions made in treatment of shock acceleration need revision. In particular, the escape of particles from SNRs may be more complex than generally assumed., Comment: 16 pages, 4 figures, accepted to PRD

Following our recent paper (Petrosian & Chen 2010), we have developed an inversion method to determine the basic characteristics of the particle acceleration mechanism directly and non-parametrically from observations under the leaky box framework. In the above paper, we demonstrated this method for obtaining the energy dependence of the escape time. Here, by converting the Fokker-Planck equation to its integral form, we derive the energy dependences of the energy diffusion coefficient and direct acceleration rate for stochastic acceleration in terms of the accelerated and escaping particle spectra. Combining the regularized inversion method of Piana et al. 2007 and our procedure, we relate the acceleration characteristics in solar flares directly to the count visibility data from RHESSI. We determine the timescales for electron escape, pitch angle scattering, energy diffusion, and direct acceleration at the loop top acceleration region for two intense solar flares based on the regularized electron flux spectral images. The X3.9 class event shows dramatically different energy dependences for the acceleration and scattering timescales, while the M2.1 class event shows a milder difference. The M2.1 class event could be consistent with the stochastic acceleration model with a very steep turbulence spectrum. A likely explanation of the X3.9 class event could be that the escape of electrons from the acceleration region is not governed by a random walk process, but instead is affected by magnetic mirroring, in which the scattering time is proportional to the escape time and has an energy dependence similar to the energy diffusion time., Comment: 11 pages, 5 figures, accepted for publication in ApJ

Where particle acceleration and plasma heating take place in relation to magnetic reconnection is a fundamental question for solar flares. We report analysis of an M7.7 flare on 2012 July 19 observed by SDO/AIA and RHESSI. Bi-directional outflows in forms of plasmoid ejections and contracting cusp-shaped loops originate between an erupting flux rope and underlying flare loops at speeds of typically 200-300 km/s up to 1050 km/s. These outflows are associated with spatially separated double coronal X-ray sources with centroid separation decreasing with energy. The highest temperature is located near the nonthermal X-ray loop-top source well below the original heights of contracting cusps near the inferred reconnection site. These observations suggest that the primary loci of particle acceleration and plasma heating are in the reconnection outflow regions, rather than the reconnection site itself. In addition, there is an initial ascent of the X-ray and EUV loop-top source prior to its recently recognized descent, which we ascribe to the interplay among multiple processes including the upward development of reconnection and the downward contractions of reconnected loops. The impulsive phase onset is delayed by 10 minutes from the start of the descent, but coincides with the rapid speed increases of the upward plasmoids, the individual loop shrinkages, and the overall loop-top descent, suggestive of an intimate relation of the energy release rate and reconnection outflow speed., Comment: Accepted by ApJ, March 04, 2013; 18 pages, 14 figures, online movies at http://lmsal.com/~weiliu/my_papers/2012-08_flare-cusp-SADs/movies.zip

We report one of several homologous non-radial eruptions from NOAA active region (AR) 11158 that are strongly modulated by the local magnetic field as observed with the Solar Dynamic Observatory (SDO). A small bipole emerged in the sunspot complex and subsequently created a quadrupolar flux system. Non-linear force-free field (NLFFF) extrapolation from vector magnetograms reveals its energetic nature: the fast-shearing bipole accumulated ~2e31 erg free energy (10% of AR total) over just one day despite its relatively small magnetic flux (5% of AR total). During the eruption, the ejected plasma followed a highly inclined trajectory, over 60 degrees with respect to the radial direction, forming a jet-like, inverted-Y shaped structure in its wake. Field extrapolation suggests complicated magnetic connectivity with a coronal null point, which is favorable of reconnection between different flux components in the quadrupolar system. Indeed, multiple pairs of flare ribbons brightened simultaneously, and coronal reconnection signatures appeared near the inferred null. Part of the magnetic setting resembles that of a blowout-type jet; the observed inverted-Y structure likely outlines the open field lines along the separatrix surface. Owing to the asymmetrical photospheric flux distribution, the confining magnetic pressure decreases much faster horizontally than upward. This special field geometry likely guided the non-radial eruption during its initial stage., Comment: 10 pages, 7 figures, 4 movies. Accepted for publication in ApJ

We report the evolution of magnetic field and its energy in NOAA active region 11158 over 5 days based on a vector magnetogram series from the Helioseismic and Magnetic Imager (HMI) on board the Solar Dynamic Observatory (SDO). Fast flux emergence and strong shearing motion led to a quadrupolar sunspot complex that produced several major eruptions, including the first X-class flare of Solar Cycle 24. Extrapolated non-linear force-free coronal fields show substantial electric current and free energy increase during early flux emergence near a low-lying sigmoidal filament with sheared kilogauss field in the filament channel. The computed magnetic free energy reaches a maximum of ~2.6e32 erg, about 50% of which is stored below 6 Mm. It decreases by ~0.3e32 erg within 1 hour of the X-class flare, which is likely an underestimation of the actual energy loss. During the flare, the photospheric field changed rapidly: horizontal field was enhanced by 28% in the core region, becoming more inclined and more parallel to the polarity inversion line. Such change is consistent with the conjectured coronal field "implosion", and is supported by the coronal loop retraction observed by the Atmospheric Imaging Assembly (AIA). The extrapolated field becomes more "compact" after the flare, with shorter loops in the core region, probably because of reconnection. The coronal field becomes slightly more sheared in the lowest layer, relaxes faster with height, and is overall less energetic., Comment: Eq. (A1) corrected

[Abridged]We present analysis of a pair of unusually energetic coronal hard X-ray (HXR) sources detected by RHESSI during the impulsive phase of an X3.9 class solar flare on 2003 November 3, which simultaneously shows two intense footpoint (FP) sources. A distinct loop top (LT) coronal source is detected up to ~150 keV and a second (upper) coronal source up to ~80 keV. These photon energies are much higher than commonly observed in coronal sources and pose grave modeling challenges. The LT source in general appears higher in altitude with increasing energy and exhibits a more limited motion compared to the expansion of the thermal loop. The high energy LT source shows an impulsive time profile and its nonthermal power law spectrum exhibits soft-hard-soft evolution during the impulsive phase, similar to the FP sources. The upper coronal source exhibits an opposite spatial gradient and a similar spectral slope compared to the LT source. These properties are consistent with the model of stochastic acceleration of electrons by plasma waves or turbulence. However, the LT and FP spectral index difference (varying from ~0-1) is much smaller than commonly measured and than that expected from a simple stochastic acceleration model. Additional confinement or trapping mechanisms of high energy electrons in the corona are required. Comprehensive modeling including both kinetic effects and the macroscopic flare structure may shed light on this behavior. These results highlight the importance of imaging spectroscopic observations of the LT and FP sources up to high energies in understanding electron acceleration in solar flares. Finally, we show that the electrons producing the upper coronal HXR source may very likely be responsible for the type III radio bursts at the decimetric/metric wavelength observed during the impulsive phase of this flare., Comment: 13 emulateapj pages, 10 figures; accepted for publication in ApJ

We present an overview of solar flares and associated phenomena, drawing upon a wide range of observational data primarily from the RHESSI era. Following an introductory discussion and overview of the status of observational capabilities, the article is split into topical sections which deal with different areas of flare phenomena (footpoints and ribbons, coronal sources, relationship to coronal mass ejections) and their interconnections. We also discuss flare soft X-ray spectroscopy and the energetics of the process. The emphasis is to describe the observations from multiple points of view, while bearing in mind the models that link them to each other and to theory. The present theoretical and observational understanding of solar flares is far from complete, so we conclude with a brief discussion of models, and a list of missing but important observations., Comment: This is an article for a monograph on the physics of solar flares, inspired by RHESSI observations. The individual articles are to appear in Space Science Reviews (2011)

The model of stochastic acceleration of particles by turbulence has been successful in explaining many observed features of solar flares. Here we demonstrate a new method to obtain the accelerated electron spectrum and important acceleration model parameters from the high resolution hard X-ray observations provided by the Reuven Ramaty High Energy Solar Spectroscopic Imager (RHESSI). In our model, electrons accelerated at or very near the loop top produce thin target bremsstrahlung emission there and then escape downward producing thick target emission at the loop footpoints. Based on the electron flux spectral images obtained by the regularized inversion of the RHESSI count visibilities, we derive several important parameters for the acceleration model. We apply this procedure to the 2003 November 03 solar flare, which shows a loop top source up to 100--150 keV in hard X-ray with a relatively flat spectrum in addition to two footpoint sources. The results imply presence of strong scattering and a high density of turbulence energy with a steep spectrum in the acceleration region., Comment: 4 pages, 2 figures; revised manuscript submitted to ApJL