Abstracts:High Fidelity (HiFi) identity customization with text-to-image generation has gained a lot of interest from all four quadrants, such as industries, consumers, researchers, and digital content creators. Such generational models are capable of personalizing images with pretrained diffusion models without extensive fine-tuning. However, existing works often compromise HiFi or generative behavior of the original model due to computational constraints associated with training identity customization on consumer electronic devices. Furthermore, when using auxiliary images for fusion, existing models often compromise the identity customization. In this regard, we propose Dynamic high-fidelity Identity Customization and Enhancement (DICE) that integrates a vision transformer (ViT), specifically dealing with facial and non-facial images to extract semantic features, a dynamic and multi-model contrastive fusion strategy, denoising diffusion model, and a composite loss function. The DICE leverages evolved feature extraction, multi-scale feature fusion, adaptive contrastive paths, and adaptive composite loss to achieve high fidelity, editability, and minimal refinement to the base model even for the fusion of base image with the auxiliary one. Such tuning-free identity customization is appropriate for the consumers on their resource constrained electronic devices, as it requires no retraining, shifting the computational burden to a one-time, server-side training process. Experiments demonstrate that DICE outperforms existing state-of-the-art methods while offering a flexible solution for personalized image generation.

Abstracts:Current deepfake detection methods primarily focus on exploring inter-frame inconsistencies using convolutional networks, neglecting the investigation of long-range spatiotemporal inconsistencies. Simultaneously, these methods rely on single-feature exploitation for forgery detection, resulting in limited generalization capability and robustness. To address these issues, this paper proposes a novel network that comprehensively utilizes illumination-geometric features and facial forgery trace features to excavate deepfake artifacts across multiple scales. The architecture comprises three main components: First, the Lighting-Geometric Information Capture Module (LGCM) integrates facial landmark normal vectors and illumination coefficients to construct comprehensive spatiotemporal representations. Then, the Bi-directional Multiscale Enhancement Module (BMEM) captures attention information between different frames in the spatial domain and models inter-frame discrepancy attention in the temporal domain. Furthermore, the Spatio-temporal Attention Module (STAM) mines global semantics and adaptively derives long-range spatiotemporal representations. Experimental results demonstrate that the proposed method achieves high AUC values on the four subsets of FF++ C40, Celeb-DF, and DFDC datasets, outperforming the comparative methods. Similarly, cross-forgery method detection validates the robustness and generalization capability of the proposed approach.

Abstracts:In recent years, the widespread application of agricultural consumer electronic devices, such as sensors, actuators, controllers, unmanned aerial vehicles (UAVs), and robots, has significantly enhanced the intelligence and efficiency of agricultural production. These devices have become indispensable tools for modern farming, enabling precise monitoring, control, and automation of various agricultural processes. However, the current technologies employed in these devices have certain limitations in terms of optimizing computational resources and reducing carbon emissions. This is a pressing issue that needs to be addressed, especially against the global backdrop of actively promoting carbon neutrality and developing low-carbon sustainable agriculture.

Abstracts:Demand forecasting and cybersecurity have become a strategic concern for the global supply chain management (SCM) of consumer electronics. This paper enhances security by utilizing an AI-driven model to optimize supply chain resilience in the consumer electronics market. This research introduces a novel deep learning technique, ResilientAI-SCM, which utilizes a Temporal Convolution Network (TCN) for demand forecasting and incorporates an innovative active cyber risk assessment model to counter rising threats. This model employs advanced deep learning time series techniques to forecast demand at the consumer order level with exceptional accuracy, even in shifting market conditions. The novelty of this research lies in its utilization of OptimX, a real-time multi-objective optimization technique that simultaneously controls inventory levels, resource allocation, and potential risks. The OptimX model has a Custom Weighted Temporal Loss (WTL), which adjusts the loss function to focus on more recent trends and increase the demand for accurate demand prediction during critical time periods. Eliminating blind spots in cyber risk analysis enhances the essential resilience of supply chain nodes against potential threats. Moreover, deep learning frameworks are described here to mitigate cyber risk, which is consolidated with supply chain node risk. A retail consumer electronics case study reveals that the model can boost demand prediction, improve inventory control in high-availability scenarios, and enhance disruption defense through control at supply chain arms. AI powers the approach and shows promise for expansion and variation to other consumer electronics markets. By combining TCNs with OptimX, the framework ensures scalable, accurate, and efficient supply chain management. This approach underscores the transformative potential of deep learning and optimization in enhancing resilience, security, and competitiveness in the consumer electronics supply market.

Abstracts:Metaverse is an emerging paradigm of modern consumer entertainment that aims to create a fully immersive, hyper-spatial, and extremely interoperable virtual space for smart applications, emphasizing a human-centric approach. Benefiting from edge computing and 5G networks, there has been a surge in the development of massive intelligent services (e.g., real-time motion sensing games) based on Convolutional Neural Networks (CNNs) deployed at the edge, which require plenty of computational and communication resources. However, the huge volume of service requests can overload edge servers, thus decreasing the quality of service (QoS). Given the limited resources available at the edge, service offloading at the cloud can be employed to ensure the QoS. In this paper, we design MP-DOB, a load balance-aware offloading strategy with CNN model partitioning, which jointly optimizes the inference delay and energy consumption in the service process. Specifically, we partition a CNN model into sub-models in sequential and spatial manners. Then, part of these sub-models is considered to be offloaded at the cloud according to the game theory-based policy. Parallelly, the other part of sub-models is re-scheduled within the edge to further balance the load of edge servers. Comparative experiments are conducted to indicate the effectiveness of our proposed MP-DOB.

Abstracts:Vehicular ad hoc networks (VANETs) are positioned to support 6G-enabled wireless energy sharing among electric vehicles, a capability that is valuable with limited charging infrastructure. Exposing energy exchange to the wireless channel introduces security and timing risks, including AI-enabled attacks that can disrupt service and safety. This study presents present cTrust, a unified framework for trust management, dynamic clustering, and energy sharing. cTrust computes per-vehicle trust from direct observations and neighbor recommendations using temporal decay and source-credibility weights; it aggregates and propagates trust locally to limit communication overhead. Clusters are formed using trust scores and a Network Efficiency Factor (NEF), while a compact request/approval protocol governs energy transactions with ledger updates for accountability. RSUs coordinate cluster maintenance, with localized computation providing continuity under intermittent connectivity. In simulations covering 50–150 vehicles over 300–500 m2 areas (DSR, 7–8 Mb/s), cTrust achieved a 98% successful transmission rate with 120 ms average latency. Detection rates for DoS, jamming, spoofing, and tampering were 95%, 94%, 96%, and 93%, respectively, with a 1.7 s end-to-end response from detection to mitigation. The framework operated at 200 W with 95% operational efficiency and remained stable under density changes and network partitioning.

Abstracts:The rapid expansion of consumer IoT systems, from smart homes to wearables, generates massive heterogeneous data streams while exposing new attack surfaces to AI-enabled threats. Traditional anomaly detection remains typically limited by unimodal learning, static models, centralized processing, and poor adaptability to concept drift. To overcome these challenges, we propose NeuroAgent-X, a self-evolving cognitive agent that integrates neuro-symbolic reasoning, attention-guided multi-modal fusion, and federated swarm intelligence. The framework features a dual-pathway encoder for structured and unstructured data, a reinforcement-driven self-evolution module with episodic and semantic memory, and selective parameter sharing for privacy-preserving distributed learning. Evaluations on SWaT, UCI HAR, and WADI datasets show superior anomaly detection (94.2% precision, 92.7% recall, 93.4% F1, 0.979 AUC) with sub-120ms latency across cloud, fog, and edge deployments. NeuroAgent-X also demonstrates resilience to concept drift (94% retention), zero-day attacks (89.2% F1), and adversarial perturbations (86.5% F1 under $\epsilon $ =0.1), outperforming state-of-the-art baselines. Furthermore, its federated design reduces communication overhead by 78% while preserving data locality. These results establish NeuroAgent-X as a scalable and robust paradigm for securing consumer IoT against adaptive, AI-enabled threats.

Abstracts:Artificial Intelligence (AI) and the Internet of Things (IoT) are transforming modern computing by enabling intelligent decision-making across distributed devices. Federated Learning (FL) facilitates decentralized model training while preserving data privacy, making it ideal for large-scale AI applications. However, FL faces a major challenge of high communication overhead due to frequent model updates between clients and the central server. Existing solutions, such as compressing gradients, struggle to balance efficiency and accuracy, particularly in resource-constrained edge environments. There is a growing need for trustworthy, communication-efficient FL solutions that ensure privacy, fairness, and security. To address this, we propose a lightweight novel Hierarchical Federated Learning (HFL) framework that integrates adaptive model pruning, quantization, and model communication and aggregation frequency optimization. First, we introduce a joint model pruning and quantization approach that dynamically adjusts pruning ratios and quantization levels, reducing communication costs while maintaining high accuracy. Second, we develop a fairness aware Stackelberg game-based model communication frequency optimization mechanism, where clients, edge servers, and the central server collaboratively determine optimal update frequencies to balance overhead and convergence speed. Third, we enhance privacy protection using Selective Homomorphic Encryption (SHE) and introduce a verifiable model trust assessment to ensure secure participation of edge devices. Extensive experiments validate our framework’s effectiveness in optimizing communication efficiency, improving model accuracy, and ensuring fast convergence under diverse FL scenarios. The results demonstrate that our approach outperforms state-of-the-art methods in terms of reducing communication overhead, enhancing privacy, and achieving robust learning across varying network and computational constraints.

Abstracts:Human activity recognition (HAR) using WiFi enables non-intrusive monitoring in consumer electronics healthcare. However, it suffers from multipath fading, device heterogeneity and scarcely labeled fine-grained data. To address these challenges, we develop a neurosymbolic artificial intelligence (AI) architecture for WiFi-based HAR. The model first segments raw WiFi channel state information (CSI) into fixed-length windows. Each window is processed by three parallel encoders, i.e., a one-dimensional convolutional neural network for local feature extraction, a Transformer with positional encoding for capturing global context, and a multilayer perceptron (MLP) that generates rule-based embeddings. The outputs of these streams are then fused via a learnable multi-head cross-attention mechanism, which amplifies salient motion cues and suppresses noise. Finally, the fused sequence is averaged over time and classified under a combined data-driven and semantic-rule loss. Four benchmark datasets, i.e., SignFi, Widar3.0, UT-HAR, and NTU-HAR are used to train and test our model. Experimental results determined that our model achieves up to 99.72% accuracy, consistently outperforming state-of-the-art (SOTA) methods. This lightweight and transparent framework supports practical deployment of explainable HAR on resource-constrained consumer electronics.