arXiv subjects

Zhifei Yu

Publications and source records attributed to Zhifei Yu.

9 recordsLinked to original sources

arXiv2025

Towards Real-World Deepfake Detection: A Diverse In-the-wild Dataset of Forgery Faces

Deepfakes, leveraging advanced AIGC (Artificial Intelligence-Generated Content) techniques, create hyper-realistic synthetic images and videos of human faces, posing a significant threat to the authenticity of social media. While this real-world threat is increasingly prevalent, existing academic evaluations and benchmarks for detecting deepfake forgery often fall short to achieve effective application for their lack of specificity, limited deepfake diversity, restricted manipulation techniques.To address these limitations, we introduce RedFace (Real-world-oriented Deepfake Face), a specialized facial deepfake dataset, comprising over 60,000 forged images and 1,000 manipulated videos derived from authentic facial features, to bridge the gap between academic evaluations and real-world necessity. Unlike prior benchmarks, which typically rely on academic methods to generate deepfakes, RedFace utilizes 9 commercial online platforms to integrate the latest deepfake technologies found "in the wild", effectively simulating real-world black-box scenarios.Moreover, RedFace's deepfakes are synthesized using bespoke algorithms, allowing it to capture diverse and evolving methods used by real-world deepfake creators. Extensive experimental results on RedFace (including cross-domain, intra-domain, and real-world social network dissemination simulations) verify the limited practicality of existing deepfake detection schemes against real-world applications. We further perform a detailed analysis of the RedFace dataset, elucidating the reason of its impact on detection performance compared to conventional datasets. Our dataset is available at: https://github.com/kikyou-220/RedFace.

cs.CV↗

arXiv2025

AI-assisted hyper-dimensional broadband quantum memory with efficiency above 90% in warm atoms

High-dimensional broadband quantum memory significantly expands quantum information processing capabilities, but the memory efficiency becomes insufficient when extended to high dimensions. We demonstrate an efficient quantum memory for hyper-dimensional photons encoded with orbital angular momentum (OAM) and spin angular momentum (SAM). OAM information is encoded from -5 to +5, combined with SAM encoding, enabling up to 22 dimensions. To ensure high memory efficiency, an artificial intelligence algorithm, a modified Differential Evolution (DE) algorithm using Chebyshev sampling, is developed to obtain a perfect signal-control waveform matching. Memory efficiency is experimentally achieved at 92% for single-mode Gaussian signal, 91% for information dimension of 6 and 80% for dimensional number to 22. The fidelity is achieved up to 99% for single-mode Gaussian signal, 95.5% for OAM information, 97.4% for SAM information, and 92% for whole hyper-dimensional signal, which is far beyond no-cloning limitation. Our results demonstrate superior performance and potential applications in high-dimensional quantum information processing. This achievement provides a crucial foundation for future quantum communication and quantum computing.

quant-ph↗

arXiv2024

NumbOD: A Spatial-Frequency Fusion Attack Against Object Detectors

With the advancement of deep learning, object detectors (ODs) with various architectures have achieved significant success in complex scenarios like autonomous driving. Previous adversarial attacks against ODs have been focused on designing customized attacks targeting their specific structures (e.g., NMS and RPN), yielding some results but simultaneously constraining their scalability. Moreover, most efforts against ODs stem from image-level attacks originally designed for classification tasks, resulting in redundant computations and disturbances in object-irrelevant areas (e.g., background). Consequently, how to design a model-agnostic efficient attack to comprehensively evaluate the vulnerabilities of ODs remains challenging and unresolved. In this paper, we propose NumbOD, a brand-new spatial-frequency fusion attack against various ODs, aimed at disrupting object detection within images. We directly leverage the features output by the OD without relying on its internal structures to craft adversarial examples. Specifically, we first design a dual-track attack target selection strategy to select high-quality bounding boxes from OD outputs for targeting. Subsequently, we employ directional perturbations to shift and compress predicted boxes and change classification results to deceive ODs. Additionally, we focus on manipulating the high-frequency components of images to confuse ODs' attention on critical objects, thereby enhancing the attack efficiency. Our extensive experiments on nine ODs and two datasets show that NumbOD achieves powerful attack performance and high stealthiness.

cs.CV↗

arXiv2023

Detecting and Corrupting Convolution-based Unlearnable Examples

Convolution-based unlearnable examples (UEs) employ class-wise multiplicative convolutional noise to training samples, severely compromising model performance. This fire-new type of UEs have successfully countered all defense mechanisms against UEs. The failure of such defenses can be attributed to the absence of norm constraints on convolutional noise, leading to severe blurring of image features. To address this, we first design an Edge Pixel-based Detector (EPD) to identify convolution-based UEs. Upon detection of them, we propose the first defense scheme against convolution-based UEs, COrrupting these samples via random matrix multiplication by employing bilinear INterpolation (COIN) such that disrupting the distribution of class-wise multiplicative noise. To evaluate the generalization of our proposed COIN, we newly design two convolution-based UEs called VUDA and HUDA to expand the scope of convolution-based UEs. Extensive experiments demonstrate the effectiveness of detection scheme EPD and that our defense COIN outperforms 11 state-of-the-art (SOTA) defenses, achieving a significant improvement on the CIFAR and ImageNet datasets.

cs.CV↗

arXiv2022

Sensitivity-enhanced magnetometry using nitrogen-vacancy ensembles via adaptively complete transitions overlapping

Nitrogen-vacancy (NV) centers in diamond are suitable sensors of high-sensitivity magnetometry which have attracted much interest in recent years. Here, we demonstrate sensitivity-enhanced ensembles magnetometry via adaptively complete transitions overlapping with a bias magnetic field equally projecting onto all existing NV orientations. Under such conditions, the spin transitions corresponding to different NV orientations are completely overlapped which will bring about an obviously improved photoluminescence contrast. And we further introduce particle swarm optimization into the calibration process to generate this bias magnetic field automatically and adaptively using computer-controlled Helmholtz coils. By applying this technique, we realize an approximate 1.5 times enhancement and reach the magnetic field sensitivity of $\rm855\ pT/\sqrt{Hz}$ for a completely overlapped transitions compared to $\rm 1.33\ nT/\sqrt{\rm Hz}$ for a separate transition on continuous-wave magnetometry. Our approach can be conveniently applied to direction-fixed magnetic sensing and obtain the potentially maximum sensitivity of ensemble-NV magnetometry.

quant-ph↗

arXiv2022

Sensing performance enhancement via asymmetric gain optimization in the atom-light hybrid interferometer

The SU (1,1)-type atom-light hybrid interferometer (SALHI) is a kind of interferometer that is sensitive to both the optical phase and atomic phase. However, the loss has been an unavoidable problem in practical applications and greatly limits the use of interferometers. Visibility is an important parameter to evaluate the sensing performance of interferometers. Here, we experimentally demonstrate the mitigating effect of the loss on visibility of the SALHI via asymmetric gain optimization, where the maximum threshold of loss to visibility close to $100\%$ is increased. Furthermore, we theoretically find that the optimal condition for the largest visibility is the same as that for the enhancement of signal-to-noise ratio (SNR) to the best value in the presence of losses using the intensity detection, indicating that the visibility can act as an experimental operational criterion for SNR improvement in practical applications. Improvement of the interference visibility means achievement of SNR enhancement. Our results provide a significant foundation for practical application of the SALHI in radar and ranging measurements.

quant-ph↗

arXiv2021

Effects of losses on the sensitivity of an actively correlated Mach-Zehnder interferometer

We theoretically studied the quantum Cram\'{e}r-Rao bound of an actively correlated Mach-Zehnder interferometer (ACMZI), where the quantum Fisher information obtained by the phase-averaging method can give the proper phase-sensing limit without any external phase reference. We numerically calculate the phase sensitivities with the method of homodyne detection and intensity detection in the presence of losses. Under lossless and very low loss conditions, the ACMZI is operated in a balanced case to beat the standard quantum limit (SQL). As the loss increases, the reduction in sensitivity increases. However within a certain range, we can adjust the gain parameters of the beam recombination process to reduce the reduction in sensitivity and realize the sensitivity can continue to beat the SQL in an unbalanced situation. Our scheme provides an optimization method of phase estimation in the presence of losses.

quant-ph↗

arXiv2019

A 9-Hr CV With One Outburst in 4 Years of Kepler Data

During a visual search through the Kepler main-field lightcurves, we have discovered a cataclysmic variable (CV) that experienced only a single 4-day long outburst over four years, rising to three times the quiescent flux. During the four years of non-outburst data the Kepler photometry of KIC 5608384 exhibits ellipsoidal light variations (`ELV') with a $\sim$12% amplitude and period of 8.7 hours. Follow-up ground-based spectral observations have yielded a high-quality radial velocity curve and the associated mass function. Additionally, H$\alpha$ emission lines were present in the spectra even though these were taken while the source was presumably in quiescence. These emission lines are at least partially eclipsed by the companion K star. We utilize the available constraints of the mass function, the ELV amplitude, Roche-lobe filling condition, and inferred radius of the K star to derive the system masses and orbital inclination angle: $M_{\rm wd} \simeq 0.46 \pm 0.02 \, M_\odot$, $M_{\rm K} \simeq 0.41 \pm 0.03 \, M_\odot$, and $i \gtrsim 70^\circ$. The value of $M_{\rm wd}$ is the lowest reported for any accreting WD in a cataclysmic variable. We have also run binary evolution models using MESA to infer the most likely parameters of the pre-cataclysmic binary. Using the mass-transfer rates from the model evolution tracks we conclude that although the rates are close to the critical value for accretion disk stability, we expect KIC 5608384 to exhibit dwarf nova outbursts. We also conclude that the accreting white dwarf most likely descended from a hot subdwarf and, most notably, that this binary is one of the first bona fide examples of a progenitor of AM CVn binaries to have evolved through the CV channel.

astro-ph.SR↗

arXiv2018

High-performance Raman quantum memory with optimal control

Quantum memories with high efficiency and fidelity are essential for long-distance quantum communication and information processing. Techniques have been developed for quantum memories based on atomic ensembles. The atomic memories relying on the atom-light resonant interaction usually suffer from the limitations of narrow bandwidth. The far-off-resonant Raman process has been considered a potential candidate for use in atomic memories with large bandwidths and high speeds. However, to date, the low memory efficiency remains an unsolved bottleneck. Here, we demonstrate a high-performance atomic Raman memory in Rb87 vapour with the development of an optimal control technique. A memory efficiency of 82.6% for 10-ns optical pulses is achieved and is the highest realized to date in atomic Raman memories. In particular, an unconditional fidelity of up to 98.0%, significantly exceeding the no-cloning limit, is obtained with the tomography reconstruction for a single-photon level coherent input. Our work marks an important advance of atomic Raman memory towards practical applications in quantum information processing.

quant-ph↗