RecipeMT, a mathematical model, is proposed to generate authentic recipe ideation across diverse culinary styles. Existing approaches to recipe design often overlook the intricate relationships between ingredient affinities and flavor pairings, which are crucial for creating recipes tailored to specific cuisines. RecipeMT addresses this limitation by leveraging flavor-sharing score weights and cascaded transformers to quantify ingredient affinities. The model not only captures the global relationships between flavor pairings and ingredient affinities across culinary styles, but also understands the interdependence between individual ingredients and the overall recipe composition. This enables RecipeMT to select the most suitable set of ingredients to ensure the generation of novel and authentic recipes. The performance of RecipeMT is evaluated using three datasets: RecipeMT_Chinese, RecipeMT_Meat, and RecipeMind. Through rigorous experiments and qualitative analysis, RecipeMT demonstrates its effectiveness in advancing recipe ideation, particularly in terms of generating recipes that are authentic to their respective culinary styles.

Pattern recognition systems are now being deployed in high-stakes environments where data integrity is critical, but often untrusted. Data-poisoning attacks are a serious threat, allowing attackers to inject malicious samples that change the behavior of the model while keeping the validation accuracy seemingly unchanged. We propose WFR-Guard, a security-motivated, model-agnostic defense framework that combines cryptographic watermarking (WM) with randomized feature verification. WFR-Guard embeds imperceptible least significant bit (LSB) watermarks into trusted data subsets to enable verifiable provenance, while epoch-wise random projection disrupts adversarial alignment without requiring changes to the classifier architecture or a separate post-hoc retraining stage. We formalize a cryptographic security game, and bound the adversary’s success probability to below 3% . Extensive experiments on CIFAR-10, GTSRB, and PathMNIST under label flipping (LF), feature collision (FC), BadNets backdoor (BD), blended trigger (BT), and adaptive mimicry (AM) attacks demonstrate that WFR-Guard maintains ≥ 89% detection accuracy (DA) even at 50% poisoning intensity, while preserving clean-data accuracy up to 98.5% . The defense achieves low FPR, strong white-box robustness, and lightweight inference overhead ( ∼ 2.3 ms per image). These results establish WFR-Guard as a practical, verifiable solution for securing high-confidence pattern recognition systems against a wide range of data-poisoning attacks. The framework is particularly relevant for safety-critical applications where corrupted training data can lead to severe operational consequences, including intelligent transportation systems and AI-assisted healthcare platforms.

Complex nonlinearities and slow dynamics are common characteristics that should be considered in industrial process modeling. However, the intrinsic physical relationships among process variables, especially the spatial and temporal correlations, have received less attention. To leverage these relationships, a novel Graph Siamese network (GraphSiamese) is proposed to extract the spatial-temporal slow features of process data based on an en coder-decoder structure of graph convolutional networks (GCNs). In GraphSiamese, the process graph is learned by canonical correlation analysis of intricate interactions between the process variables. Then, the GCN is sub stituted for the autoencoder of a traditional Siamese network, focusing on capturing slow features of industrial processes with a graph structure. The network is trained in an end-to-end manner to simultaneously reconstruct the original inputs and the adjacency matrix, capture slow features, and predict quality variables. By visualizing the graph-structured slow features across both spatial-temporal fusion aspects, the proposed method enhances interpretability of a soft sensing model, moving beyond traditional black-box models. The effectiveness of the GraphSiamese method is validated through a public sulfur recovery unit benchmark and a real CO2 absorption column case, demonstrating its robustness and accuracy in industrial process quality prediction modeling.

Statistical heterogeneity and personalization remain key challenges in Federated Learning (FL). We propose FedDWl2, a novel FL method that combines dynamic weight aggregation with L2 regularization to address these issues. Specifically, the dynamic weight aggregation module adaptively adjusts the weights of global and local models according to each client’s data distribution, effectively capturing client-specific feature information and alleviating the impact of data heterogeneity (e.g., feature space shift). Meanwhile, L2 regularization is integrated to mitigate overfitting, enhance the generalization ability of the global model, and further assist in personalized model optimization by balancing the consistency between the global model and local models. FedDWl2 dynamically adjusts the weights of global and local models based on each client’s data distribution and employs L2 regularization to mitigate overfitting. Experiments on benchmark datasets such as CIFAR-10 and MNIST demonstrate that FedDWl2 achieves an average improvement of 0.5

Supervised learning that rely exclusively on class labels often suffer from insufficient guidance for training deep neural networks (DNNs), resulting in inherent model biases and limited interpretability. To address these challenges, this study introduces the Consistent Interpretable Model (CIM) for interpretable image classification. The proposed CIM is constructed based on three core components. First, causal inference analysis is employed to model the learning process of DNNs, establishing a theoretical foundation for identifying spurious feature-task correlations. Second, the unbiasedness of task features is ensured by aligning original image embeddings with task-relevant prior knowledge. Finally, rich semantic knowledge from contrastively pre-trained text encoders is exploited to strengthen the alignment between the model’s task feature distribution and real-world scenarios. Extensive experiments show that CIM substantially improves DNN performance, particularly in fine-grained and out-of-distribution classification (achieving over 8% improvement in Top-1 accuracy) tasks. Moreover, CIM outperforms existing methods in terms of causality (with over a 10% gain in Grad-CAM’s Overall metric) and consistency (with over a 4% gain in Grad-CAM’s CH metric) of model explanations. Additionally, the alignment of prompt text representations with task object representations in CIM enables textual explanations for DNN decisions. The code is publicly available at https://github.com/DLAIResearch/CIM.

Deep Learning (DL) has significantly advanced artificial intelligence (AI) across numerous applications. However, its foundational reliance on centralized data collection introduces critical limitations concerning privacy, security, and scalability. With the continued proliferation of the Internet of Things (IoT), massive volumes of sensitive data are generated at the network edge, necessitating collaborative learning systems capable of securely sharing information without compromising confidentiality. Federated Learning (FL) partially addresses these challenges by enabling on-device model training, but its dependence on a central coordinator remains a core vulnerability, creating communication bottlenecks, fairness constraints, and susceptibility to single-point-of-failure (SPOF) risks. Swarm Learning (SL), developed in collaboration with Hewlett Packard Enterprise (HPE), represents a decentralized advancement that mitigates these limitations by eliminating central orchestration. Through the integration of blockchain technology with distributed machine learning (ML), SL establishes a secure, transparent, and fault-tolerant paradigm for peer-to-peer model exchange and aggregation. This survey provides a comprehensive overview of the architectural foundations, enabling technologies, and applications of SL in sensitive domains, including healthcare, the Internet of Vehicles (IoV), and Industrial IoT. In addition, it introduces a comparative taxonomy that systematically categorizes existing research by scope, methodological approach, and evaluation maturity. The review concludes by identifying key research directions—such as lightweight consensus protocols, energy-efficient optimization, and cross-chain interoperability—to advance the development of practical, secure, and privacy-preserving SL systems.

Artificial intelligence (AI) is transforming dairy quality assessment by enabling rapid, non-destructive, and data-driven monitoring across the dairy supply chain. Conventional analytical methods, although accurate, are often labor-intensive, time-consuming, and unsuitable for real-time applications. The integration of advanced sensing technologies with AI has therefore emerged as a promising approach for improving dairy quality control and food safety. This review examines recent advances in AI-enabled dairy quality assessment, covering spectroscopy-based sensing technologies, biomimetic sensing systems, machine vision, and multimodal data fusion. Particular attention is given to their applications in major dairy products, including raw milk, milk powder, cheese, and yogurt. Comparative analysis shows that AI-assisted sensing significantly enhances the accuracy, efficiency, and automation of compositional analysis, adulteration detection, microbial screening, freshness evaluation, and defect inspection, while multimodal approaches offer new opportunities for comprehensive quality assessment. The review further highlights the emerging role of explainable artificial intelligence (XAI) in improving the transparency and trustworthiness of dairy detection systems. Recent advances in feature attribution, decision interpretation, and uncertainty quantification are discussed, together with challenges related to data scarcity, model generalization, cross-instrument transferability, and explainability evaluation. Future developments in multimodal foundation models, standardized datasets, and explainable intelligent systems are expected to accelerate the deployment of trustworthy and scalable dairy quality assurance frameworks.

Blockchain can preserve supply-chain records, but ledger integrity alone does not show whether a participant should be trusted in a future risk-sensitive transaction. Existing reputation systems mainly address product evidence, global feedback aggregation, or review authenticity, while giving less attention to repeated bilateral inflation, identity multiplicity, and unfair decay for honest participants with sparse histories. We present , a participant trust framework that uses blockchain as an evidence and provenance layer rather than as the source of trust. Governed interaction outcomes are converted into bounded evidence. Repeated interactions between the same pair are discounted, low counterparty diversity is penalized, governance-supplied identity confidence weights positive evidence, and scores decay toward a neutral prior according to verified interaction volume. Identity, contract, outcome, and update provenance remain on chain, while nonlinear reputation computation is performed off chain and checked on chain for admissibility. In controlled simulations with 30 seeded runs and matched interaction traces, the full model reduces mean collusive gain to 0.1443, compared with 0.3688 for naive mean evidence and 0.3585 for static decay. With ten identities under one controller, the reputation inflation ratio falls to 0.8723, while three comparison baselines remain above 1.08. On identical newcomer traces, volume-aware decay increases mean newcomer reputation from 0.6626 to 0.7589 and reduces the false low-trust rate from 0.3633 to 0.1683. Paired analysis confirms these improvements across runs. The results support a bounded reduction in reputation distortion, not attacker detection. Deployment evaluation and calibration with operational data are still required before production use.

Cell-cultured meat and fermented protein, as novel food ingredients, involve complex biotechnological methods and metabolic regulation in their biological manufacturing processes. However, traditional manufacturing methods face challenges in formula optimization, structure-activity relationship prediction, and analysis of complex metabolites. This article reviews the application progress of artificial intelligence (AI) and machine learning in the production of cell-cultured meat and fermented protein. Firstly, it elaborates on the key technical principles of machine learning and deep learning; secondly, from three dimensions of upstream strain design and component function prediction, midstream biological process formula optimization and monitoring, and downstream product safety assessment and quality control, it systematically analyzes how AI, through computational technology, achieves protein structure prediction, medium formula optimization and process monitoring, and metabolic characteristic analysis; finally, it discusses challenges such as data standardization, model interpretability, and regulatory compliance, and looks forward to the future development directions of automated laboratories and digital twin ecosystems. This article aims to provide theoretical references for the intelligent upgrade of the design of novel food ingredients and the biological manufacturing process enabled by AI.

To address the challenges of aflatoxin M1 (AFM1) detection in the dairy industry, this study proposes a neural network model named STANet, based on a spatiotemporal attention mechanism, for probabilistic prediction of exceeding AFM1 in raw milk. This study utilizes nearly 1 million raw milk spatiotemporal monitoring data from December 2022 to October 2024 of a Chinese dairy company, generates high-quality synthetic data to alleviate the sample imbalance problem by the TABSYN method, and effectively integrates the spatiotemporal information with the basic features by using the dynamic weight fusion mechanism and attention pooling process. The experimental results show that STANet significantly outperforms other models in terms of accuracy (0.835), especially in the identification of sparse positive samples, with F1 scores of 0.842 and 0.828 for category 0 and category 1, respectively. The ablation experiments further validate the spatiotemporal features and dynamic weighting mechanism to enhance the performance of the model. Feature importance analysis reveals the significant contribution of features to the prediction. This study provides a data-driven framework to support the early probabilistic detection and monitoring of AFM1 exceedance, which has the potential to reduce screening costs in large-scale raw milk quality control. Future improvements may be achieved by incorporating biochemical validation experiments and refining data pre-processing strategies.

As a core component of critical infrastructure, industrial control systems (ICS) generate large volumes of finegrained operational data. Distributed artificial intelligence offers opportunities for secure collaboration across ICS clients, but devices such as programmable logic controllers (PLCs), remote terminal units (RTUs), and supervisory control and data acquisition (SCADA) systems remain limited to basic control and monitoring, lacking capabilities for intelligent analysis. Privacy concerns further hinder effective data utilization. To address these challenges, we propose a distributed, reliability-enhanced collaborative training framework, termed ICS physical equipment-Edge Server-Blockchain (IEEB), which integrates federated learning, edge computing, and block-chain. Edge servers enable localized model training under resource constraints, while blockchain and smart contracts provide decentralized management, process automation, dynamic incentives, and anti-poisoning mechanisms. IEEB ensures auditability, traceability, and secure collaboration among heterogeneous ICS. Implemented with Hyperledger Fabric and evaluated on five types of PLCs, IEEB improved average AUC by 9.1% on the SWaT, WADI, and MSL datasets, reduced training time by 64.7% and batches by 80.3%, and maintained robustness against up to 20% malicious clients.

Class imbalance remains a major challenge in real-world classification tasks. To address this, we propose Diffusion-Enhanced Ensemble Learning (DEEL), a unified framework that synergistically integrates diffusion-based data augmentation and multi-objective ensemble optimization for binary classification tasks. Specifically, we design a Dynamic Attention Diffusion Model (DADM) to generate diverse and realistic minority class samples through a forward noise and reverse denoising process. By incorporating temporal embeddings, residual connections, and attention mechanisms, DADM enhances the fidelity and distributional alignment of the generated data. Complementing this, an ensemble learning strategy based on the Non-dominated Sorting Genetic Algorithm II (NSGA-II) optimizes the fusion of multiple base classifiers across F1-score, G-mean, and AUC metrics. Extensive experiments on 26 real-world imbalanced datasets demonstrate that DEEL improves average F1-score and G-mean by 21.7 % and 24.8 %, respectively, over competitive baselines. Moreover, visualization and Jensen-Shannon distance analyses quantitatively verify the high diversity and distributional coherence of DADM-generated samples, underscoring their effectiveness for imbalanced learning.

The architecture of Industrial Control Systems (ICS) has evolved into an integrated cyber-physical system, introducing covert, zero-dynamic cyberattack vectors that threaten the reliability of ICS. We address two issues: (1) how to use decentralized solutions to defend resource-constrained ICS against highly covert zero-dynamic attacks; (2) how to improve performance (meeting real-time, high-frequency interaction demands) while enhancing reliability via decentralization. We propose a defense model integrating blockchain, zero-knowledge proofs, and smart contract (SC) obfuscation to bolster ICS resilience. A customized zk-SNARK algorithm enables efficient identity authentication, completed in under 3 ms. The Garble framework obfuscates SCs, concealing ICS device IP addresses and disrupting attack chains. A blockchain acts as a secure intermediary between the engineer workstation (EW) and programmable logic controller (PLC). To reduce blockchain overhead, we refine a proportional-integral-derivative (PID)-based roulette wheel algorithm. Obfuscated SCs resist decompilation by tools such as Objdump, ensuring robust protection. Our node selection mechanism balances security and diversity, mitigating systemic biases like the Matthew effect. By leveraging blockchain to supply computational power for encrypting and protecting ICS data flows, we offer new insights into defending ICS against highly covert cyberattacks. Experimental evaluation validates the model's effectiveness under real-world ICS scenarios.

Intrusion Detection Systems (IDS) powered by Federated Learning (FL) are vital for securing Internet of Medical Things (IoMT) environments while preserving data privacy. However, FL-IDS remain highly vulnerable to poison ing and backdoor attacks, where adversaries manipulate local updates to compromise global model integrity. A key challenge arises from inconsistent convergence due to gradient magnitude disparities, leading to conflicting update directions, gradient forgetting, and covert backdoor persistence. Existing defenses, which rely on auxiliary validation data or client exclusion, are insufficient to effectively address these issues. To address these vulnerabilities, we propose a novel FL-based IDS framework underpinned by an enhanced Deep Neural Network (DNN) architecture. The novelty of this work lies in jointly addressing gradient incon sistency and adversarial robustness through a unified, data-independent mechanism that stabilizes learning dynamics and mitigates adversarial effects. Optimized for deployment on local clients, the framework provides accurate threat detection and adapts to heterogeneous network environments. The proposed defense intro duces three key contributions: (1) a gradient-stabilization strategy combining Intra-Layer Directional Gradient Projection and Gradient-Magnitude Normalization across layers, aligning updates to mitigate backdoor risks, (2) selective parameterization through gradient-based importance scoring, enabling efficient low-dimensional repre sentations while retaining critical learning gradients, and (3) an Isolation Forest-driven mechanism to identify and exclude poisoned gradients, thereby preserving model integrity. The proposed framework demonstrates strong ro bustness under both IID and non-IID settings, effectively handling data heterogeneity. Experimental results show that, in the intrusion detection task, the framework improves average accuracy by 10.18%. Furthermore, when applied to a cross-domain vision-based setting for defending against poisoning attacks in image classification, it achieves an average accuracy improvement of 6.89% compared to state-of-the-art federated defense methods.

Data imbalance is one of the main reasons for the degradation of machine learning model performance. In supervised learning, class imbalance often leads to a shift in the decision boundary, causing minority class samples containing critical information to be misclassified. To address this issue, this study proposes an oversampling method based on Generative Adversarial Networks (GANs) to improve the classification performance on imbalanced data. We design a novel training loss function for generator to better capture the distributional characteristics of input samples, thereby generating higher-quality samples to construct a balanced training dataset. Subsequently, we train and test the model on multiple datasets. Experimental results show that, compared to existing oversampling methods, the proposed approach significantly improves the performance of classification algorithms on 20 real-world imbalanced datasets.

The cybersecurity threats targeting industrial control systems (ICSs) are evolving with increasing sophistication. Addressing the detection blind spots in existing source code analysis techniques, this study reveals a dual security paradox arising from code sensitivity: privacy leakage risks caused by decompilation techniques and integrity verification deficiencies in reverse engineering. This article investigates three critical challenges: 1) what are the component flow process and detection elements of ICS component source code? 2) how can high-performance and reliable tracing and traceability be provided for ICS component source code exceptions and routine detection? and 3) how can privacy enhancement and trusted detection of ICS component source code with high sensitivity be achieved? This article proposes a blockchain-integrated trusted detection framework for ICS (BCTD-ICS), delivering groundbreaking solutions: 1) establishing a lifecycle circulation model that systematically maps component types, stakeholders, and detection parameters; 2) developing a tripartite collaborative architecture [blockchain- identification resolution zero-knowledge proofs (ZKPs)], featuring a traceability mechanism with trusted identification codes (resolution efficiency: 40 ms/105 queries) to eliminate decompilation-induced privacy risks; and 3) creating an industrial-oriented privacy enhancement system utilizing DBSCAN clustering for intelligent sampling (26% compression rate on BCN3D Moveo) and optimizing ZK-SNARK protocols through Shamir's secret sharing, establishing a backdoor-resistant distributed parameter generation system (time delay increment < 100 ms). Experimentally verified, our solution enables ICS component code detection supply-chain-wise without sensitive data leakage in real-world industries. This work establishes a novel trusted detection paradigm for ICS, advancing detection efficiency and credibility under strict privacy preservation requirements, meeting Industry 4.0 security demands.

Powerful facial image manipulation offered by encoder-based GAN inversion techniques raises concerns about potential misuse in identity fraud and misinformation. This study introduces the Multi-Domain and Multi-Dimensional Feature Fusion (MDFusion) method, a novel approach that counters encoder-based GAN inversion by generating adversarial samples. Firstly, MDFusion transforms the luminance channel of the target image into spatial, frequency, and spatial-frequency hybrid domains. Secondly, we use the specifically adapted Feature Pyramid Network (FPN) to extract and fuse high-dimensional and low-dimensional features that enhance the robustness of adversarial noise generation. Then, we embed adversarial noise into the spatial-frequency hybrid domain to produce effective adversarial samples. Finally, the adversarial samples are guided by our designed hybrid training loss to achieve a balance between imperceptibility and effectiveness. Tests were conducted on five encoder-based GAN inversion models using ASR, LPIPS, and FID metrics. These tests demonstrated the superiority of MDFusion over 13 baseline methods, highlighting its robust defense and generalization abilities. The implementation code is available at https://github.com/LuckAlex/MDFusion.

The Internet of Medical Things (IoMT) has transformed healthcare by enabling seamless communication among medical devices, supporting real-time monitoring, diagnostics, vital patient data tracking, improved patient care, disease prediction, early warning, and enhanced operational efficiency. Due to the sensitive nature of healthrelated data, the adoption of IoMT has raised significant privacy and security concerns, prompting comprehensive evaluation of IoMT security and making it a prime target for cyberattacks such as ransomware, denial-ofservice (DoS) attacks, and malware. Securing IoMT requires efficient data processing across distributed systems to ensure both confidentiality and availability. Parallel and distributed computing can address scalability and performance challenges in IoMT security, particularly in enabling real-time monitoring and threat detection across multiple interconnected devices. This survey conducts a systematic literature review (SLR) of IoMT security to analyze key issues, categorize security threats, attack vectors, and vulnerabilities, and examine how emerging technologies such as blockchain, machine learning (ML), and physically unclonable functions (PUF) are strengthening IoMT security. This SLR reviews IoMT security research published between 2020 and 2024 to identify challenges and provide insights for future researchers and developers of new IoMT security models. This SLR offers practitioners and researchers guidance for developing reliable and resilient IoMT security systems in the decentralized healthcare industry.

This review explores the application of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in food safety detection and risk prediction. This paper highlights the advantages of CNNs in image processing and feature recognition, as well as the powerful capabilities of RNNs (especially their variant LSTM) in time series data modeling. This paper also makes a comparative analysis in many aspects: Firstly, the advantages and disadvantages of traditional food safety detection and risk prediction methods are compared with deep learning technologies such as CNNs and RNNs. Secondly, the similarities and differences between CNNs and fully connected neural networks in processing image data are analyzed. Furthermore, the advantages and disadvantages of RNNs and traditional statistical modeling methods in processing time series data are discussed. Finally, the application directions of CNNs in food safety detection and RNNs in food safety risk prediction are compared. This paper also discusses combining these deep learning models with technologies such as the Internet of Things (IoT), blockchain, and federated learning to improve the accuracy and efficiency of food safety detection and risk warning. Finally, this paper mentions the limitations of RNNs and CNNs in the field of food safety, as well as the challenges in the interpretability of the model, and suggests the use of interpretable artificial intelligence (XAI) technology to improve the transparency of the model.