arXiv subjects

Zhaoyu Chen

Publications and source records attributed to Zhaoyu Chen.

66 records · Page 4Linked to original sources

arXiv2023

Explicit and Implicit Knowledge Distillation via Unlabeled Data

Data-free knowledge distillation is a challenging model lightweight task for scenarios in which the original dataset is not available. Previous methods require a lot of extra computational costs to update one or more generators and their naive imitate-learning lead to lower distillation efficiency. Based on these observations, we first propose an efficient unlabeled sample selection method to replace high computational generators and focus on improving the training efficiency of the selected samples. Then, a class-dropping mechanism is designed to suppress the label noise caused by the data domain shifts. Finally, we propose a distillation method that incorporates explicit features and implicit structured relations to improve the effect of distillation. Experimental results show that our method can quickly converge and obtain higher accuracy than other state-of-the-art methods.

cs.CV↗

arXiv2022

Towards Practical Certifiable Patch Defense with Vision Transformer

Patch attacks, one of the most threatening forms of physical attack in adversarial examples, can lead networks to induce misclassification by modifying pixels arbitrarily in a continuous region. Certifiable patch defense can guarantee robustness that the classifier is not affected by patch attacks. Existing certifiable patch defenses sacrifice the clean accuracy of classifiers and only obtain a low certified accuracy on toy datasets. Furthermore, the clean and certified accuracy of these methods is still significantly lower than the accuracy of normal classification networks, which limits their application in practice. To move towards a practical certifiable patch defense, we introduce Vision Transformer (ViT) into the framework of Derandomized Smoothing (DS). Specifically, we propose a progressive smoothed image modeling task to train Vision Transformer, which can capture the more discriminable local context of an image while preserving the global semantic information. For efficient inference and deployment in the real world, we innovatively reconstruct the global self-attention structure of the original ViT into isolated band unit self-attention. On ImageNet, under 2% area patch attacks our method achieves 41.70% certified accuracy, a nearly 1-fold increase over the previous best method (26.00%). Simultaneously, our method achieves 78.58% clean accuracy, which is quite close to the normal ResNet-101 accuracy. Extensive experiments show that our method obtains state-of-the-art clean and certified accuracy with inferring efficiently on CIFAR-10 and ImageNet.

cs.CV↗

arXiv2022

Efficient universal shuffle attack for visual object tracking

Recently, adversarial attacks have been applied in visual object tracking to deceive deep trackers by injecting imperceptible perturbations into video frames. However, previous work only generates the video-specific perturbations, which restricts its application scenarios. In addition, existing attacks are difficult to implement in reality due to the real-time of tracking and the re-initialization mechanism. To address these issues, we propose an offline universal adversarial attack called Efficient Universal Shuffle Attack. It takes only one perturbation to cause the tracker malfunction on all videos. To improve the computational efficiency and attack performance, we propose a greedy gradient strategy and a triple loss to efficiently capture and attack model-specific feature representations through the gradients. Experimental results show that EUSA can significantly reduce the performance of state-of-the-art trackers on OTB2015 and VOT2018.

cs.CV↗

arXiv2021

CMUA-Watermark: A Cross-Model Universal Adversarial Watermark for Combating Deepfakes

Malicious applications of deepfakes (i.e., technologies generating target facial attributes or entire faces from facial images) have posed a huge threat to individuals' reputation and security. To mitigate these threats, recent studies have proposed adversarial watermarks to combat deepfake models, leading them to generate distorted outputs. Despite achieving impressive results, these adversarial watermarks have low image-level and model-level transferability, meaning that they can protect only one facial image from one specific deepfake model. To address these issues, we propose a novel solution that can generate a Cross-Model Universal Adversarial Watermark (CMUA-Watermark), protecting a large number of facial images from multiple deepfake models. Specifically, we begin by proposing a cross-model universal attack pipeline that attacks multiple deepfake models iteratively. Then, we design a two-level perturbation fusion strategy to alleviate the conflict between the adversarial watermarks generated by different facial images and models. Moreover, we address the key problem in cross-model optimization with a heuristic approach to automatically find the suitable attack step sizes for different models, further weakening the model-level conflict. Finally, we introduce a more reasonable and comprehensive evaluation method to fully test the proposed method and compare it with existing ones. Extensive experimental results demonstrate that the proposed CMUA-Watermark can effectively distort the fake facial images generated by multiple deepfake models while achieving a better performance than existing methods.

cs.CV↗

arXiv2021

RPATTACK: Refined Patch Attack on General Object Detectors

Nowadays, general object detectors like YOLO and Faster R-CNN as well as their variants are widely exploited in many applications. Many works have revealed that these detectors are extremely vulnerable to adversarial patch attacks. The perturbed regions generated by previous patch-based attack works on object detectors are very large which are not necessary for attacking and perceptible for human eyes. To generate much less but more efficient perturbation, we propose a novel patch-based method for attacking general object detectors. Firstly, we propose a patch selection and refining scheme to find the pixels which have the greatest importance for attack and remove the inconsequential perturbations gradually. Then, for a stable ensemble attack, we balance the gradients of detectors to avoid over-optimizing one of them during the training phase. Our RPAttack can achieve an amazing missed detection rate of 100% for both Yolo v4 and Faster R-CNN while only modifies 0.32% pixels on VOC 2007 test set. Our code is available at https://github.com/VDIGPKU/RPAttack.

cs.CV↗

arXiv2010

Phase Diagram of Pressure-induced Superconductivity and its Relation to Hall Coefficient in Bi2Te3 Single Crystal

Pressure-induced superconductivity and its relation to corresponding Hall coefficient (RH) have been reported for Bi2Te3, one of known topological insulators. A full phase diagram is presented which shows a complex dependence of the superconducting transition temperature as a function of pressure over an extensive range. High-pressure RH measurements reveal a close relation of these complex behaviors, particularly, a dramatic change of dRH/dP before structural phase transition and a pressure-induced crossover on RH in the high pressure phase were observed.

cond-mat.supr-con↗

arXiv2010

Eight new quasars discovered by LAMOST in one extragalactic field

We report the discovery of eight new quasars in one extragalactic field (five degree centered at RA=$08^h58^m08.2^s$, Dec=$01^o32'29.7''$) with the LAMOST commissioning observations on December 18, 2009. These quasars, with $i$ magnitudes from 16.44 to 19.34 and redshifts from 0.898 to 2.773, were not identified in the SDSS spectroscopic survey, though six of them with redshifts less than 2.5 were selected as quasar targets in SDSS. Except one source without near-IR $Y$-band data, seven of these eight new quasars meet a newly proposed quasar selection criterion involving both near-IR and optical colors. Two of them were found in the 'redshift desert' for quasars ($z$ from 2.2 to 3), indicating that the new criterion is efficient for recovering the missing quasars with similar optical colors as stars. Although LAMOST met some problems during the commissioning observations, we were still able to identify other 38 known SDSS quasars in this field, with $i$ magnitudes from 16.24 to 19.10 and redshifts from 0.297 to 4.512. Our identifications imply that a substantial fraction of quasars may be missing in the previous quasar surveys. The implication of our results to the future LAMOST quasar survey is discussed.

astro-ph.CO↗

arXiv2010

Unusual Electronic Structure and Observation of Dispersion Kink in CeFeAsO Parent Compound of FeAs-Based Superconductors

We report the first comprehensive high-resolution angle-resolved photoemission measurements on CeFeAsO, a parent compound of FeAs-based high temperature superconductors with a mangetic/structural transition at $\sim$150 K. In the magnetic ordering state, four hole-like Fermi surface sheets are observed near $Γ$(0,0) and the Fermi surface near M(+/-$π$,+/-$π$) shows a tiny electron-like pocket at M surrounded by four Dirac cone-like strong spots. The unusual Fermi surface topology deviates strongly from the band structure calculations. The electronic signature of the magnetic/structural transition shows up in the dramatic change of the quasiparticle scattering rate. A dispersion kink at $\sim$ 25meV is for the first time observed in the parent compound of Fe-based superconductors.

cond-mat.supr-con↗

arXiv2010

A very bright i=16.44 quasar in the `redshift desert' discovered by LAMOST

The redshift range from 2.2 to 3, is known as the 'redshift desert' of quasars because quasars with redshift in this range have similar optical colors as normal stars and are thus difficult to be found in optical sky surveys. A quasar candidate, SDSS J085543.40-001517.7, which was selected by a recently proposed criterion involving near-IR $Y-K$ and optical $g-z$ colors, was identified spectroscopically as a new quasar with redshift of 2.427 by the LAMOST commissioning observation in December 2009 and confirmed by the observation made with the NAOC/Xinglong 2.16m telescope in March 2010. This quasar was not targeted in the SDSS spectroscopic survey because it locates in the stellar locus of the optical color-color diagrams, while it is clearly separated from stars in the $Y-K$ vs. $g-z$ diagram. Comparing with other SDSS quasars we found this new quasar with $i$ magnitude of 16.44 is apparently the brightest one in the redshift range from 2.3 to 2.7. From the spectral properties we derived its central black hole mass as $(1.4\sim3.9) \times 10^{10} M_\odot$ and the bolometric luminosity as $3.7\times 10^{48}$ \ergs, which indicates that this new quasar is intrinsically very bright and belongs to the most luminous quasars in the universe. Our identification supports that quasars in the redshift desert can be found by the quasar selection criterion involving the near-IR colors. More missing quasars are expected to be recovered by the future LAMOST spectroscopic surveys, which is important to the study of the cosmological evolution of quasars at redshift higher than 2.2.

astro-ph.CO↗

arXiv2009

SDSS spectroscopy for blazars in the $Fermi$ LAT bright AGN sample

We have collected all available spectra and photometric data from SDSS catalogue for bright AGNs complied from the first three months of the Fermi large area telescope all-sky survey. Based on the 106 high-confidence and 11 low-confidence association bright AGN list, the photometry data are collected from SDSS DR7 for 28 sources (12 BL Lacs and 16 FSRQs), two of which are low-confidence association bright AGNs. Among these 28 SDSS photometric sources, SDSS spectra are available for 20 sources (6 BL Lacs and 14 FSRQs). The black hole mass M_BH and the broad line region (BLR) luminosity were obtained for 14 FSRQs by measuring the line-width and strength of broad emission lines from SDSS spectra. The broad emission lines measurements of five FSRQs are presented for the first time in this work. The optical continuum emission of these 14 FSRQs is found to be likely dominated by the nonthermal jet emission through comparing the relationship between the broad Mg II line and continuum luminosity to that of radio quiet AGNs. The black hole mass of 14 FSRQs ranges from 10^8.2 to 10^9.9 solar mass, with most of sources larger than 10^9 solar mass. The Eddington ratio L_bol/L_Edd ranges from 10^-1.5 to ~ 1. This implies that the optically thin, geometrically thick accretion disk may exist in these FSRQs.

astro-ph.GA↗

arXiv2009

M_BH - sigma relation in SDSS flat-spectrum radio quasars

The relationship between the black hole mass and velocity dispersion indicated with [O III] line width is investigated for a sample of 87 flat-spectrum radio quasars (FSRQs) selected from SDSS DR3 quasar catalogue. We found the M_bh - sigma_[O III] relation is deviated from Tremaine et al. relation for nearby inactive galaxies, with a larger black hole mass at given velocity dispersion. There is no strong evidence of cosmology evolution in M_bh - sigma_[O III] relation up to z~0.8. A significant correlation between the [O III] luminosity and Broad Line Region (BLR) luminosity is found. When transferring the [O III] luminosity to Narrow Line Region (NLR) luminosity, the BLR luminosity is, on average, larger than NLR one by about one order of magnitude. We found a strong correlation between the synchrotron peak luminosity and NLR luminosity, which implies a tight relation between the jet physics and accretion process.

astro-ph.GA↗

arXiv2009

Flat-Spectrum Radio Quasars from SDSS DR3 Quasar Catalogue

We constructed a sample of 185 Flat Spectrum Radio Quasars (FSRQs) by cross-correlating the Shen et al.'s SDSS DR3 X-ray quasar sample with FIRST and GB6 radio catalogues. From the spectrum energy distribution (SED) constructed using multi-band (radio, UV, optical, Infrared and X-ray) data, we derived the synchrotron peak frequency and peak luminosity. The black hole mass and the broad line region (BLR) luminosity (then the bolometric luminosity) were obtained by measuring the line-width and strength of broad emission lines from SDSS spectra. We define a subsample of 118 FSRQs, of which the nonthermal jet emission is thought to be dominated over the thermal emission from accretion disk and host galaxy. For this subsample, we found 25 FSRQs having synchrotron peak frequency > 10^{15} Hz, which is higher than the typical value for FSRQs. While only a weak anti-correlation is found between the synchrotron peak frequency and peak luminosity, it becomes significant when combining with the Wu et al.'s sample of 170 BL Lac objects. At similar peak frequency, the peak luminosity of FSRQs with $\nupeak > 10^{15}$ Hz is systematically higher than that of BL Lac objects, with some FSRQs out of the range covered by BL Lac objects. Although high $\nupeak$ are found in some FSRQs, they do not reach the extreme value of BL Lacs. For the subsample of 118 FSRQs, we found significant correlations between the peak luminosity and black hole mass, the Eddington ratio, and the BLR luminosity, indicating that the jet physics may be tightly related with the accretion process.

astro-ph.GA↗