Search Results (14)

Deploying neural networks on encrypted data requires efficient evaluation of nonlinear activations, especially the ReLU function, without decryption. While the CKKS homomorphic encryption scheme supports packed arithmetic over approximate numbers efficiently, its approximate semantics make direct nonlinear evaluation difficult, and polynomial surrogates often introduce approximation error and non-discrete outputs. In this work, we present a task-specific, non-interactive construction for discrete ReLU evaluation in CKKS by combining modulus-switch-based discretization with interpolation-driven lookup-table (LUT) evaluation. We instantiate this design in two complementary schemes. The first uses trigonometric Hermite interpolation and functional bootstrapping to compute a discrete sign indicator, which is then combined with the encrypted input through conditional multiplication to obtain the ReLU output; this variant is compact and suitable for lightweight settings. The second uses iterative most-significant-bit (MSB) bootstrapping to support larger plaintext moduli and higher-precision regimes through repeated digit extraction. A common enabler of both schemes is a discretization step that maps approximate CKKS plaintexts to a finite integer representation; exactness in our setting therefore refers to exact evaluation over this discretized representation, while the deviation from the original CKKS plaintext is governed by the discretization error analyzed in Lemma 1. Experiments on encrypted MNIST inference and the accompanying LUT/storage analysis indicate that the proposed schemes preserve competitive accuracy relative to polynomial-approximation baselines while maintaining manageable auxiliary storage under the reported parameter settings. These results suggest that interpolation-based discrete activation is a promising alternative to polynomial approximation for selected CKKS-based encrypted inference tasks. Full article

Federated learning (FL) enables collaborative model training across distributed devices and robotic systems while preserving data privacy, making it well-suited for swarm robotics and edge-device-powered intelligence. However, FL remains vulnerable to adversarial behaviors such as data and model poisoning, particularly in real-world deployments where detection methods must operate under strict computational and communication constraints. This paper presents a practical, real-world federated learning framework that enhances robustness to adversarial agents in a ROS2-based edge-device swarm environment. The proposed system integrates the Federated Averaging (FedAvg) algorithm with a lightweight average cosine similarity-based filtering method to detect and suppress harmful model updates during aggregation. Unlike prior work that primarily evaluates poisoning defenses in simulated environments, this framework is implemented and evaluated on physical hardware, consisting of a laptop-based aggregator and multiple Raspberry Pi worker nodes. A convolutional neural network (CNN) based on the MobileNetV3-Small architecture is trained on the MNIST dataset, with one worker executing a sign-flipping model poisoning attack. Experimental results show that FedAvg alone fails to maintain meaningful model accuracy under adversarial conditions, resulting in near-random classification performance with a final global model accuracy of 11% and a loss of 2.3. In contrast, the integration of cosine similarity filtering demonstrates effective detection of sign-flipping model poisoning in the evaluated ROS2 swarm experiment, allowing the global model to maintain model accuracy of around 90% and loss around 0.37, which is close to baseline accuracy of 93% of the FedAvg algorithm only under no attack with a very minimal increase in loss, despite the presence of an attacker. The proposed method also maintains a false positive rate (FPR) of around 0.01 and a false negative rate (FNR) of around 0.10 of the global model in the presence of an attacker, which is a minimal difference from the baseline FedAvg-only results of around 0.008 for FPR and 0.07 for FNR. Additionally, the proposed method of FedAvg + cosine similarity filtering maintains computational statistics similar to baseline FedAvg with no attacker. Baseline results show an average runtime of about 34 min, while our proposed method shows an average runtime of about 35 min. Also, the average size of the global model being shared among workers remains consistent at around 7.15 megabytes, showing little to no increase in message payload sizes between baseline results and our proposed method. These results demonstrate that computationally lightweight cosine similarity-based detection methods can be effectively deployed in real-world, resource-constrained robotic swarm environments, providing a practical path toward improving robustness in real-world federated learning deployments beyond simulation-based evaluation. Full article

Fully homomorphic encryption (FHE) offers a promising solution for privacy-preserving machine learning by enabling arbitrary computations on encrypted data. However, the efficient evaluation of non-linear functions—such as the ReLU activation function over large integers—remains a major obstacle in practical deployments, primarily due to high bootstrapping overhead and limited precision support in existing schemes. In this paper, we propose $\mathsf{LargeIntReLU}$, a novel framework that enables efficient homomorphic ReLU evaluation over large integers (7–11 bits) via full-domain bootstrapping. Central to our approach is a signed digit decomposition algorithm, $\mathsf{SignedDecomp}$, that partitions a large integer ciphertext into signed 6-bit segments using three new low-level primitives: $\mathsf{LeftShift}$, $\mathsf{HomMod}$, and $\mathsf{CipherClean}$. This decomposition preserves arithmetic consistency, avoids cross-segment carry propagation, and allows parallelized bootstrapping. By segmenting the large integer and processing each chunk independently with optimized small-integer bootstrapping, we achieve homomorphic ReLU with full-domain bootstrapping, which significantly reduces the total number of sequential bootstrapping operations required. The security of our scheme is guaranteed by TFHE. Experimental results demonstrate that the proposed method reduces the bootstrapping cost by an average of 28.58% compared to state-of-the-art approaches while maintaining 95.2% accuracy. With execution times ranging from 1.16 s to 1.62 s across 7–11 bit integers, our work bridges a critical gap toward a scalable and efficient homomorphic ReLU function, which is useful in privacy-preserving machine learning. Furthermore, an end-to-end encrypted inference test on a CNN model with the MNIST dataset confirms its practicality, achieving 88.85% accuracy and demonstrating a complete pipeline for privacy-preserving neural network evaluation. Full article

With the increasing sophistication of Artificial Intelligence (AI), traditional digital steganography methods face a growing risk of being detected and compromised. Adversarial attacks, in particular, pose a significant threat to the security and robustness of hidden information. To address these challenges, this paper proposes a novel AI-based steganography framework designed to enhance the security of concealed messages within digital images. Our approach introduces a multi-stage embedding process that utilizes a sequence of encoder models, including a base encoder, a residual encoder, and a dense encoder, to create a more complex and secure hiding environment. To further improve robustness, we integrate Wavelet Transforms with various deep learning architectures, namely Convolutional Neural Networks (CNNs), Bayesian Neural Networks (BNNs), and Graph Convolutional Networks (GCNs). We conducted a comprehensive set of experiments on the FashionMNIST and MNIST datasets to evaluate our framework’s performance against several adversarial attacks. The results demonstrate that our multi-stage approach significantly enhances resilience. Notably, while CNN architectures provide the highest baseline accuracy, BNNs exhibit superior intrinsic robustness against gradient-based attacks. For instance, under the Fast Gradient Sign Method (FGSM) attack on the MNIST dataset, our BNN-based models maintained an accuracy of over 98%, whereas the performance of comparable CNN models dropped sharply to between 10% and 18%. This research provides a robust and effective method for developing next-generation secure steganography systems. Full article

Quantum machine learning (QML) has attracted significant research attention over the last decade. Multiple models have been developed to demonstrate the practical applications of the quantum properties. In this study, we first demonstrate that the previously proposed quanvolutional neural network (QuanvNN) using a randomly generated quantum circuit improves the image classification accuracy of a fully connected neural network against the Modified National Institute of Standards and Technology (MNIST) dataset and the Canadian Institute for Advanced Research 10 class (CIFAR-10) dataset from 92.0% to 93.0% and from 30.5% to 34.9%, respectively. We then propose a new model referred to as a Neural Network with Quantum Entanglement (NNQE) using a strongly entangled quantum circuit combined with Hadamard gates. The new model further improves the image classification accuracy of MNIST and CIFAR-10 to 93.8% and 36.0%, respectively. Unlike other QML methods, the proposed method does not require optimization of the parameters inside the quantum circuits; hence, it requires only limited use of the quantum circuit. Given the small number of qubits and relatively shallow depth of the proposed quantum circuit, the proposed method is well suited for implementation in noisy intermediate-scale quantum computers. While promising results were obtained by the proposed method when applied to the MNIST and CIFAR-10 datasets, a test against a more complicated German Traffic Sign Recognition Benchmark (GTSRB) dataset degraded the image classification accuracy from 82.2% to 73.4%. The exact causes of the performance improvement and degradation are currently an open question, prompting further research on the understanding and design of suitable quantum circuits for image classification neural networks for colored and complex data. Full article

Deep neural networks (DNNs) have famously been applied in various ordinary duties. However, DNNs are sensitive to adversarial attacks which, by adding imperceptible perturbation samples to an original image, can easily alter the output. In state-of-the-art white-box attack methods, perturbation samples can successfully fool DNNs through the network gradient. In addition, they generate perturbation samples by only considering the sign information of the gradient and by dropping the magnitude. Accordingly, gradients of different magnitudes may adopt the same sign to construct perturbation samples, resulting in inefficiency. Unfortunately, it is often impractical to acquire the gradient in real-world scenarios. Consequently, we propose a self-adaptive approximated-gradient-simulation method for black-box adversarial attacks (SAGM) to generate efficient perturbation samples. Our proposed method uses knowledge-based differential evolution to simulate gradients and the self-adaptive momentum gradient to generate adversarial samples. To estimate the efficiency of the proposed SAGM, a series of experiments were carried out on two datasets, namely MNIST and CIFAR-10. Compared to state-of-the-art attack techniques, our proposed method can quickly and efficiently search for perturbation samples to misclassify the original samples. The results reveal that the SAGM is an effective and efficient technique for generating perturbation samples. Full article

The Convolutional Neural Network (CNN) has demonstrated excellent performance in image recognition and has brought new opportunities for sign language recognition. However, the features undergo many nonlinear transformations while performing the convolutional operation and the traditional CNN models are insufficient in dealing with the correlation between images. In American Sign Language (ASL) recognition, J and Z with moving gestures bring recognition challenges. This paper proposes a novel Two-Stream Mixed (TSM) method with feature extraction and fusion operation to improve the correlation of feature expression between two time-consecutive images for the dynamic gestures. The proposed TSM-CNN system is composed of preprocessing, the TSM block, and CNN classifiers. Two consecutive images in the dynamic gesture are used as inputs of streams, and resizing, transformation, and augmentation are carried out in the preprocessing stage. The fusion feature map obtained by addition and concatenation in the TSM block is used as inputs of the classifiers. Finally, a classifier classifies images. The TSM-CNN model with the highest performance scores depending on three concatenation methods is selected as the definitive recognition model for ASL recognition. We design 4 CNN models with TSM: TSM-LeNet, TSM-AlexNet, TSM-ResNet18, and TSM-ResNet50. The experimental results show that the CNN models with the TSM are better than models without TSM. The TSM-ResNet50 has the best accuracy of 97.57% for MNIST and ASL datasets and is able to be applied to a RGB image sensing system for hearing-impaired people. Full article

Image denoising has been a knotty issue in the computer vision field, although the developing deep learning technology has brought remarkable improvements in image denoising. Denoising networks based on deep learning technology still face some problems, such as in their accuracy and robustness. This paper constructs a robust denoising network based on a generative adversarial network (GAN). Since the neural network has the phenomena of gradient dispersion and feature disappearance, the global residual is added to the autoencoder in the generator network, to extract and learn the features of the input image, so as to ensure the stability of the network. On this basis, we proposed an optimization algorithm (OA), to train and optimize the mean and variance of noise on each node of the generator. Then the robustness of the denoising network was improved through back propagation. Experimental results showed that the model’s denoising effect is remarkable. The accuracy of the proposed model was over 99% in the MNIST data set and over 90% in the CIFAR10 data set. The peak signal to noise ratio (PSNR) and structural similarity (SSIM) values of the proposed model were better than the state-of-the-art models in the BDS500 data set. Moreover, an anti-interference test of the model showed that the defense capacities of both the fast gradient sign method (FGSM) and project gradient descent (PGD) attacks were significantly improved, with PSNR and SSIM values decreased by less than 2%. Full article

Adversarial attacks deceive deep neural network models by adding imperceptibly small but well-designed attack data to the model input. Those attacks cause serious problems. Various defense methods have been provided to defend against those attacks by: (1) providing adversarial training according to specific attacks; (2) denoising the input data; (3) preprocessing the input data; and (4) adding noise to various layers of models. Here we provide a simple but effective Noise-Fusion Method (NFM) to defend adversarial attacks against DNN image classification models. Without knowing any details about attacks or models, NFM not only adds noise to the model input at run time, but also to the training data at training time. Two l_∞-attacks, the Fast Gradient Signed Method (FGSM) and the Projected Gradient Descent (PGD), and one l₁-attack, the Sparse L1 Descent (SLD), are applied to evaluate defense effects of the NFM on various deep neural network models which used MNIST and CIFAR-10 datasets. Various amplitude noises with different statistical distribution are applied to show the defense effects of the NFM in different noise. The NFM also compares with an adversarial training method on MNIST and CIFAR-10 datasets. Results show that adding noise to the input images and the training images not only defends against all three adversarial attacks but also improves robustness of corresponding models. The results indicate possibly generalized defense effects of the NFM which can extend to other adversarial attacks. It also shows potential application of the NFM to models not only with image input but also with voice or audio input. Full article

Nowadays, hand gestures have become a booming area for researchers to work on. In communication, hand gestures play an important role so that humans can communicate through this. So, for accurate communication, it is necessary to capture the real meaning behind any hand gesture so that an appropriate response can be sent back. The correct prediction of gestures is a priority for meaningful communication, which will also enhance human–computer interactions. So, there are several techniques, classifiers, and methods available to improve this gesture recognition. In this research, analysis was conducted on some of the most popular classification techniques such as Naïve Bayes, K-Nearest Neighbor (KNN), random forest, XGBoost, Support vector classifier (SVC), logistic regression, Stochastic Gradient Descent Classifier (SGDC), and Convolution Neural Networks (CNN). By performing an analysis and comparative study on classifiers for gesture recognition, we found that the sign language MNIST dataset and random forest outperform traditional machine-learning classifiers, such as SVC, SGDC, KNN, Naïve Bayes, XG Boost, and logistic regression, predicting more accurate results. Still, the best results were obtained by the CNN algorithm. Full article

Capsule networks are a type of neural network that use the spatial relationship between features to classify images. By capturing the poses and relative positions between features, this network is better able to recognize affine transformation and surpass traditional convolutional neural networks (CNNs) when handling translation, rotation, and scaling. The stacked capsule autoencoder (SCAE) is a state-of-the-art capsule network that encodes an image in capsules which each contain poses of features and their correlations. The encoded contents are then input into the downstream classifier to predict the image categories. Existing research has mainly focused on the security of capsule networks with dynamic routing or expectation maximization (EM) routing, while little attention has been given to the security and robustness of SCAEs. In this paper, we propose an evasion attack against SCAEs. After a perturbation is generated based on the output of the object capsules in the model, it is added to an image to reduce the contribution of the object capsules related to the original category of the image so that the perturbed image will be misclassified. We evaluate the attack using an image classification experiment on the Mixed National Institute of Standards and Technology Database (MNIST), Fashion-MNIST, and German Traffic Sign Recognition Benchmark (GTSRB) datasets, and the average attack success rate can reach 98.6%. The experimental results indicate that the attack can achieve high success rates and stealthiness. This finding confirms that the SCAE has a security vulnerability that allows for the generation of adversarial samples. Our work seeks to highlight the threat of this attack and focus attention on SCAE’s security. Full article

This paper presents an approach to design convolutional neural network architectures, using the particle swarm optimization algorithm. The adjustment of the hyper-parameters and finding the optimal network architecture of convolutional neural networks represents an important challenge. Network performance and achieving efficient learning models for a particular problem depends on setting hyper-parameter values and this implies exploring a huge and complex search space. The use of heuristic-based searches supports these types of problems; therefore, the main contribution of this research work is to apply the PSO algorithm to find the optimal parameters of the convolutional neural networks which include the number of convolutional layers, the filter size used in the convolutional process, the number of convolutional filters, and the batch size. This work describes two optimization approaches; the first, the parameters obtained by PSO are kept under the same conditions in each convolutional layer, and the objective function evaluated by PSO is given by the classification rate; in the second, the PSO generates different parameters per layer, and the objective function is composed of the recognition rate in conjunction with the Akaike information criterion, the latter helps to find the best network performance but with the minimum parameters. The optimized architectures are implemented in three study cases of sign language databases, in which are included the Mexican Sign Language alphabet, the American Sign Language MNIST, and the American Sign Language alphabet. According to the results, the proposed methodologies achieved favorable results with a recognition rate higher than 99%, showing competitive results compared to other state-of-the-art approaches. Full article

Domain adaptation is a sub-field of transfer learning that aims at bridging the dissimilarity gap between different domains by transferring and re-using the knowledge obtained in the source domain to the target domain. Many methods have been proposed to resolve this problem, using techniques such as generative adversarial networks (GAN), but the complexity of such methods makes it hard to use them in different problems, as fine-tuning such networks is usually a time-consuming task. In this paper, we propose a method for unsupervised domain adaptation that is both simple and effective. Our model (referred to as TripNet) harnesses the idea of a discriminator and Linear Discriminant Analysis (LDA) to push the encoder to generate domain-invariant features that are category-informative. At the same time, pseudo-labelling is used for the target data to train the classifier and to bring the same classes from both domains together. We evaluate TripNet against several existing, state-of-the-art methods on three image classification tasks: Digit classification (MNIST, SVHN, and USPC datasets), object recognition (Office31 dataset), and traffic sign recognition (GTSRB and Synthetic Signs datasets). Our experimental results demonstrate that (i) TripNet beats almost all existing methods (having a similar simple model like it) on all of these tasks; and (ii) for models that are significantly more complex (or hard to train) than TripNet, it even beats their performance in some cases. Hence, the results confirm the effectiveness of using TripNet for unsupervised domain adaptation in image classification. Full article

Deep-learning convolutional neural networks (CNNs) have proven to be successful in various cognitive applications with a multilayer structure. The high computational energy and time requirements hinder the practical application of CNNs; hence, the realization of a highly energy-efficient and fast-learning neural network has aroused interest. In this work, we address the computing-resource-saving problem by developing a deep model, termed the Gabor convolutional neural network (Gabor CNN), which incorporates highly expression-efficient Gabor kernels into CNNs. In order to effectively imitate the structural characteristics of traditional weight kernels, we improve upon the traditional Gabor filters, having stronger frequency and orientation representations. In addition, we propose a procedure to train Gabor CNNs, termed the fast training method (FTM). In FTM, we design a new training method based on the multipopulation genetic algorithm (MPGA) and evaluation structure to optimize improved Gabor kernels, but train the rest of the Gabor CNN parameters with back-propagation. The training of improved Gabor kernels with MPGA is much more energy-efficient with less samples and iterations. Simple tasks, like character recognition on the Mixed National Institute of Standards and Technology database (MNIST), traffic sign recognition on the German Traffic Sign Recognition Benchmark (GTSRB), and face detection on the Olivetti Research Laboratory database (ORL), are implemented using LeNet architecture. The experimental result of the Gabor CNN and MPGA training method shows a 17–19% reduction in computational energy and time and an 18–21% reduction in storage requirements with a less than 1% accuracy decrease. We eliminated a significant fraction of the computation-hungry components in the training process by incorporating highly expression-efficient Gabor kernels into CNNs. Full article