Fast and Robust Copy-Move Forgery Detection Using BRIEF, FAST, and SIFT Feature Matching

The advancement of generative models has made it easier to create highly realistic Deepfake videos. This accessibility has led to a surge in research on Deepfake detection to mitigate potential misuse. Typically, Deepfake detection models utilize binary backbones, even though the training dataset contains additional exploitable information, such as the Deepfake generation method employed for each video. However, recent findings suggest that inferring a binary class from a multi-class backbone yields superior performance compared to directly employing a binary backbone. Building upon this research, our article introduces two novel methods to infer a binary class from a multi-class backbone. The first method, named root dummies , leverages the dummy triplet loss, which employs fixed vectors (i.e., dummies) instead of mined positives and negatives in the triplet loss. By training the multi-class backbone with these dummies, we can easily infer a binary class during testing by adjusting the number of dummies (from six during training to two during inference). Through this approach, we achieve an accuracy improvement of 0.23% compared to the existing inference method, without requiring additional training. The second proposed method is transfer learning. It involves training a classifier, such as a support vector machine, to predict binary classes based on the image embeddings generated by the multi-class backbone. Although this method necessitates additional training, it further enhances the model’s performance, resulting in an accuracy increase of 1.79%. In summary, our proposed methods improve the accuracy of Deepfake detection by simply modifying the number of classes during training, making them suitable for integration into a variety of existing Deepfake training pipelines. Additionally, to foster reproducible research, we have made the source code of our solution publicly available at https://github.com/beuve/DmyT .

As artificial intelligence technology advances, the generalization performance of audio deepfake detection systems has improved rapidly and significantly. However, these systems still face substantial challenges when confronted with adversarial attacks. To address this issue, many studies have proposed defense methods against such attacks. Nevertheless, only a limited number of studies have focused on defenses in the context of audio deepfake detection. With the development of deep learning, deepfake detection methods are increasingly designed and deployed based on deep learning models. As a result, deepfake detection systems can be applied across various fields and are not restricted to specific data types or domains. Such approaches are referred to as ‘domain-independent’ defense methods and can be applied to a wide range of tasks, including adversarial training. However, they have limitations, such as requiring significant computing resources or potentially degrading original performance. In this paper, we propose Inference-Time Noise Addition (ITNA), an effective and efficient adversarial attack defense method that minimizes the original performance degradation. ITNA adds Gaussian noise once to the input audio sample during the inference phase. It is highly resource efficient, as it does not require additional training or separate AI components. Furthermore, ITNA minimizes performance degradation as much as possible and provides scalability that can be used with other defense methods. We provide a theoretical explanation of the effectiveness of ITNA and validate it through experimental results. Our experiments were conducted on the ASVspoof2021 DF evaluation (official deepfake audio dataset from the ASVspoof community), In-The-Wild (collected in real environments) and DSD-Corpus datasets (includes various synthesizer samples). And experiments results demonstrate the effectiveness of ITNA, reducing the average misclassification rate and the attack success rate of detection systems by 4.48% and 31.90%, respectively. We also demonstrate the effectiveness of ITNA through comparative experiments with alternative noise type and existing noise-based defense method.

Morphological processing of derived words develops simultaneously with reading acquisition. However, the reader’s engagement in morphological segmentation may depend on the language morphological richness and orthographic transparency, and the readers’ reading skills. The current study tested the common idea that morphological segmentation is enhanced in non-transparent orthographies to compensate for the absence of phonological information. Hebrew’s rich morphology and the dual version of the Hebrew script (with and without diacritic marks) provides an opportunity to study the interaction of orthographic transparency and morphological segmentation on the development of reading skills in a within-language design. Hebrew speaking 2nd (N = 27) and 5th (N = 29) grade children read aloud 96 noun words. Half of the words were simple mono-morphemic words and half were bi-morphemic derivations composed of a productive root and a morphemic pattern. In each list half of the words were presented in the transparent version of the script (with diacritic marks), and half in the non-transparent version (without diacritic marks). Our results show that in both groups, derived bi-morphemic words were identified more accurately than mono-morphemic words, but only for the transparent, pointed, script. For the un-pointed script the reverse was found, namely, that bi-morphemic words were read less accurately than mono-morphemic words, especially in second grade. Second grade children also read mono-morphemic words faster than bi-morphemic words. Finally, correlations with a standardized measure of morphological awareness were found only for second grade children, and only in bi-morphemic words. These results, showing greater morphological effects in second grade compared to fifth grade children suggest that for children raised in a language with a rich morphology, common and easily segmented morphemic units may be more beneficial for younger compared to older readers. Moreover, in contrast to the common hypothesis, our results show that morphemic segmentation does not compensate for the missing phonological information in a non-transparent orthography, but rather that morphological segmentation is most beneficial in the highly transparent script. These results are consistent with the idea that morphological and phonological segmentation processes occur simultaneously and do not constitute alternative pathways to visual word recognition.

DeepFake has accomplished notable advancement with the AI-leveraged production and manipulation techniques of fictitious human facial images. Despite many benign and fun applications, the generated fake images can negatively influence the authenticity of online information by originating deception, manipulation, persecution, and seduction, defying societal quality and human rights, which becomes critical security and privacy threat in social networks. Hence, real-time DeepFake detection and limitation technologies on the mobile platform are essential to building a controlled, harmless DeepFake ecosystem. This paper presents a real-time, cloudless, lightweight mobile app for human visual DeepFake detection using machine learning technologies named MobiDeep (Mobile DeepFake Detection through Machine Learning-based Corneal-Specular Backseat-tering). MobiDeep stems from a hypothesis that the existing DeepFake creation methods, including replacement, editing, and synthesis, lack the ensemble with the reflective objects. Focusing on the most reflective area of a human face, corneal-specular backscatter images of eyes, we seek the similarity and consistency with multiple surrounding environment features, including color components, shapes, and textures. We have implemented a cross-platform mobile application to evaluate the performance using various input parameters and lightweight Deep Neural Network (DNN) architectures. The empirical results show that MobiDeep achieves high accuracy (98.7%) and rapid detection speed (less than 200 ms) in detecting sophisticated DeepFake images within a subsecond.

Introduction: Accurate short-term wind field prediction is essential for meteorology, aviation, and renewable energy applications. Traditional numerical weather prediction (NWP) models are limited by high computational costs and low temporal resolution, making them unsuitable for real-time forecasting. To overcome these limitations, deep learning offers a promising alternative by effectively modeling complex spatiotemporal dependencies in environmental data. Objectives: This study aims to develop a deep learning-based approach for high-fidelity short-term wind field prediction. The key goals are to: (1) capture fine-scale wind structures with high spatial and temporal accuracy, (2) reduce computational complexity compared to NWP models, and (3) evaluate the model’s performance using standard metrics. Methods: We propose a multi-layer Convolutional Long Short-Term Memory (ConvLSTM) network for wind nowcasting. ConvLSTM combines convolutional operations with temporal memory, making it ideal for spatially and temporally coherent predictions. The model is enhanced with batch normalization and dropout to prevent overfitting. It is trained on sequences of wind velocity fields (U and V components), learning intricate flow patterns across time. Results: The ConvLSTM model delivers strong predictive performance. It achieves an MSE of 0.0429, RMSE of 0.02071, and MAE of 0.0603, indicating low error. High PSNR (66.2592) and SSIM (0.9978) scores confirm the model’s ability to preserve spatial detail and structural integrity in predicted wind fields. These results demonstrate its capability to accurately capture fine-scale wind dynamics. Conclusions: The proposed ConvLSTM framework presents a reliable and efficient solution for short-term wind field prediction. It offers substantial improvements in spatial accuracy and computational speed over traditional NWP models. With high fidelity and structural consistency, this deep learning model shows strong potential for real-time wind nowcasting in critical applications.

This paper presents a calculation model for tuyere blast volume distribution in a blast furnace, integrating mechanistic simulation with data-driven approaches. By employing computational fluid dynamics (CFD), a detailed model of the blast furnace air supply system is developed, enabling precise determination of resistance coefficients and blast density under varying conditions. Building on the mechanistic foundation, a data-driven model is formulated to rapidly calculate blast volume distribution, raceway depth, and kinetic energy, achieving significant improvements in both prediction accuracy and computational speed. The model also incorporates multiple factors, including tuyere tip diameter, length, wear, coal injection rate, and oxygen enrichment rate. It reveals critical insights: increasing tuyere tip diameter boosts blast flow rate and raceway depth, while changes in tuyere length have minimal impact. Deeper tuyere insertion reduces the central dead zone but enlarges the peripheral one. Furthermore, the model highlights the trade-offs between enhanced reaction efficiency through increased coal injection and oxygen enrichment, and the resulting increase in resistance that decreases volume flow. Tuyere blockage leads to uneven blast volume distribution, with unblocked tuyeres experiencing higher blast volume, kinetic energy, and raceway depth. The key innovation of this model lies in its hybrid approach, seamlessly integrating mechanistic and data-driven techniques to offer a more accurate, real-time, and comprehensive understanding of blast furnace operations.

A direct covariance algorithm is presented far handling problems of component tolerance analysis and random input variations with a particular emphasis for utilization in computer-aided statistical electronic circuit design. It is shown that this result is applicable to a wide range of electronic circuit arrays having non-linear components. Moreover, a systematic procedure is developed for predicting in advance the expected accuracy. Numerical results comparing the direct covariance algorithm with up to 1000 Monte Carlo ensemble-averaged computer runs are provided. Contrary to popular belief, errors of 10 to 25% are obtained by using 25 to 100 Monte Carlo runs. Improvements in both accuracy and computational speed clearly demonstrate that the direct covariance algorithm is a versatile and effective computer-aided design tool.

The possibilities of the optimization of histogram of oriented gradients calculations for solving image content recognition problems described based on the 48 × 48 pixels size image example. The algorithm doesn’t change in regards to the input data and suits for the histogram of the oriented gradients calculation based on any image. The algorithm idea is taken from the work of Soojin Kim and Kyeongsoon Cho [1], which is the modification of original HOG descriptor algorithm presented by Navneet Dalal and Bill Triggs [3] aimed at optimization of the calculation speed without loosing accuracy during image content recognition using HOG descriptor to generate the set of features of the image content. The algorithm is described in detail in the next sequence of actions. 1) Original HOG feature calculation. We use it as the first step since the algorithm is the optimized version of the original HOG feature calculation. 2) Solving the aliasing problem and accuracy improvement by using the interpolation technique during the HOG feature calculation process. We use one of the normalization schemes, applying interpolation as the next calculation step. 3) Solving the redundant operations and calculation speed problems by using cell-based operations and applying from one to four described types to the cells, based on which depends the detection window cell calculation. The types are applied to cell based on the blocks intersection containing the cell. The computer program has been developed according to the selected optimized algorithm of HOG feature calculation. It was used during the image content features description and learning process and in the further computer vision research. The software implementation of the algorithm takes into account the capabilities of modern computer technology, Javascript programming language and modern needs of the image content recognition calculation speed and accuracy. The implementation of computer program logic is shown in the script examples, utilizing modularity and parallel calculation as the strong sides of Javascript, further improving HOG feature calculation speed.

<div class="htmlview paragraph">By dividing the combustion process into several phases with phase optimized skeletal mechanisms (POSM), gains in calculation speed were realized with virtually no loss in accuracy.</div> <div class="htmlview paragraph">A skeletal mechanism is a reduced mechanism where only the significant species, determined through a set of parameters (one for each species), remain with respect to a detailed mechanism. The parameter is based on a combination of sensitivity and flow analysis. Within the POSM method machine learning algorithms are used to automatically determine and recognize the major phases. Reduction is achieved by keeping only the significant species with respect to each phase. Each phase has a different mechanism, derived from the original and each is smaller than the original.</div> <div class="htmlview paragraph">This novel approach was implemented into a two-zone zero-dimensional stochastic reactor model (SRM) for an SI-Engine with a 125 species heptane-toluene mechanism and showed that calculation speed was increased by a factor 3 with almost no loss in accuracy.</div> <div class="htmlview paragraph">Further time savings and accuracy improvements from these preliminary results are expected from further reduction of the phase mechanisms</div>

In this paper, we propose an adaptive grey predictor (AGP) that aims at the short-term fading (STF) of noisy Rayleigh channels in the mobile communication systems. Here, we use a moving window to reduce the degradation by the additive white Gaussian noise and further improve the precision of the STF prediction of the adaptive segmental grey prediction model, GM (1, 1, α, β). Computer simulation results show that choosing an appropriate length of moving windows in the proposed window AGP (WAGP) can obtain better accuracy and large performance improvement on the noisy STF prediction.

A number of proposed optical flow estimation algorithms for image motion analysis use only a few subsequent frames from an image sequence to perform calculation. Others, that use many frames, as is case in frequency and energy based, as well as in stochastic optical flow algorithms, are better suited for off-line work. We present our approach in adopting a classic gradient optical flow algorithm for on-line work over extended image sequence. Our experiments on the usual optical flow testing image sequences, consisting of few dozens of image frames show significant improvements in accuracy of the estimated velocity flow fields and in speed of computation. This is achieved by modification and integration of three basic processing phases: preparation of a limited spatio-temporal volume, estimation of velocity flow fields and post filtering of the flow fields. In the preparation phase we use efficient spatio-temporal filters for image smoothing. In the estimation phase a modified iterative solution that propagates results from previous calculation steps is used. Linear and nonlinear post filtering additionally improve optical flow estimates. The whole process is less sensitive to the aperture problem and produces more accurate flow field estimates than the original algorithm. It better copes with motion discontinuities as well. We evidence our achievements by a number of experiments on extended image sequences.

Auscultation, the technique of listening to heart sounds, remains a primary detection tool for diagnosing heart valve disorders. Other techniques, e.g. electrocardiography, ultrasound, etc, are accurate as well as informative; but expensive. With major advancement in the speed of computers, heart sound signals can be processed with ease by memory efficient digital signal processing and pattern recognition algorithms. This paper presents a digital signal processor (DSP) implementation of such a technique by using new threshold criteria. The proposed work can detect whether a heart sound recording belongs to a person suffering from valvular heart disease or not by giving ‘diseased’ or ‘not diseased’ decisions. The algorithm is tested for nine commonly occurring pathological problems and normal heart sound. The robustness of the algorithm is also checked against synthetically injected additive white Gaussian noise (AWGN) with different SNR levels. It is found to give an accuracy of 96.67% up to SNR values of 15 dB and 93.33% up to SNR values of 5 dB.

Incremental modeling and its application for driver fatigue estimation

Mathematics education faces global challenges that affect numeracy learning based on the results from the Program for International Student Assessment (PISA) and Trends in Mathematics and Science Study (TIMSS), that many learners struggle to master basic skills essential for both academic and everyday tasks, as stated in the study of (Smith, 2020). In seeking innovative solutions, this study assessed the mathematical competency gaps and disruptive interests of learners at Palo 19 National High School. Also, evaluated the effectiveness and acceptability of the developed NumEra-Z Program, a computer-based numeracy intervention designed to enhance accuracy and speed in mathematical tasks. Grounded in Experiential Learning Theory and Constructivist principles, the program utilized the ADDIE instructional design model to facilitate the development and evaluation of NumEra-Z, addressing the gaps in numeracy skills. There were sixty-eight (68) learners from Grades 7 to 10 who participated in the study during the second quarter of School Year 2024-2025. The research identified the least-learned competencies, mostly with fractions, basic operations, and geometric representation. Through diagnostic assessments and surveys, the disruptive interests of the learners were moderately affected by excessive peer socialization, followed by a lack of interest in abstract concepts, and then by technology misuse. Pretest and posttest results showed improvements in numeracy accuracy and speed, while a survey assessed the program's acceptability in terms of relevance, organization, engagement, and overall effectiveness. Findings revealed significant improvements in learners' numeracy accuracy and speed following participation in the NumEra-Z Program. Additionally, a positive relationship was observed between program acceptability and numeracy skills. The study highlights the potential of learning interventions like the NumEra-Z Program to address persistent gaps in numeracy proficiency while fostering learner engagement.

Thermally induced errors critically limit precision in machine tools (MT), affecting dimensional accuracy in high-precision manufacturing. Physics-based models offer a sustainable solution and can do without training measurements, unlike data-driven models, but are often unsuitable for real-time thermal error compensation due to computational demands and boundary condition uncertainties. This paper presents an enhanced ensemble Kalman filter approach to accurately and efficiently reconstruct the thermal state from limited sensor data. Utilizing a reduced finite element model generated via Krylov modal subspace techniques, the proposed method integrates real-time sensor measurements with an ensemble of perturbed models to robustly estimate thermal states and mechanical deformations. Validation experiments on a thermal test bench with 10 out of 40 temperature sensors observed demonstrate significant improvements in accuracy achieving an RMSE of $$1.5\,^{\circ }$$ 1.5 ∘ C. The computational speed is over 45 $$\times $$ × faster than real-time allowing the implementation of the digital twin for any MT. The reduction of the volumetric thermal error by consistently more than 70% confirm the potential of this digital twin-based approach for real-time thermal error compensation in precision manufacturing.