The dynamic nature of tasks and environments in Internet of Things (IoT) systems require deep learning models to continuously retrain on evolving data to ensure their effectiveness. Existing continual learning (CL) methods aim to mitigate catastrophic forgetting, where the model loses knowledge of previous tasks when learning new ones. However, these methods often ignore the memory resource competition caused by the parallel execution of multiple applications, which limits the realworld IoT application of CL in resource-constrained edge devices. In this article, we propose EDCL, a novel approach that enhances the training efficiency and model accuracy of CL methods while ensuring the uninterrupted operation of high-priority inference programs. Specifically, we first implement a custom batch sampler that can dynamically load batches and measure the memory usage and training time recorded via offline profiling. In the online stage, by monitoring the resource consumption of high-priority programs, EDCL can dynamically select batch policies that meet resource constraints and facilitate efficient training. Additionally, we propose an adaptive hierarchical buffer swap method to enhance the model’s ability to retain previously learned knowledge and mitigate forgetting. Extensive experiments show that EDCL effectively balances training efficiency and model accuracy while preventing high-priority inference programs from failing due to memory contention, demonstrating promising performance compared to baselines.

End-to-end semantic communication paradigms demonstrate substantial potential in reducing network load and compressing data redundancy. However, their inherent openness introduces significant security risks, such as unauthorized access that enables attackers to camouflage themselves among legitimate users. Moreover, legitimate users may exploit input-output data pairs to conduct model stealing attacks. Existing defense strategies generally lack user access control mechanisms and fail to provide targeted countermeasures against model inversion attacks from internal users. To address this gap, we propose KGSC-SAT, a Key-Gated Semantic Communication framework enhanced by Steganography Adversarial Training for Secure Transmission. The framework employs a key-based feature modulation method to identify authorized users, while adversarial steganography training facilitates deep feature-level masking. Experimental results demonstrate that KGSC-SAT effectively mitigates both unauthorized access and insider model inversion threats, while delivering reliable communication performance.

Multimodal Large Language Models (MLLMs), while powerful, pose significant privacy risks by memorizing and potentially exposing sensitive information linked to individuals' visual appearances. Existing machine unlearning techniques, developed primarily for text-based models, are ill-equipped to handle the deeply entangled nature of visual and semantic knowledge. To address this challenge, we introduce GAIA-UL, a novel three-stage framework that performs Surgical Unlearning of visual knowledge. Our approach first conducts a Causal Hotspot Diagnosis, using gradient-based analysis to precisely identify influential parameters within the visual-semantic pathway. Second, it performs a Targeted Adapter Intervention, surgically injecting lightweight, trainable adapters only at these hotspots while freezing the base model. Finally, it employs Semantically Orthogonal Fine-tuning, a novel objective that forces the model's internal representation of a target face to become orthogonal to embeddings of associated sensitive concepts, thereby erasing the link at a deep representational level. Extensive experiments on the MLLMU-Bench benchmark demonstrate that GAIA-UL significantly outperforms existing baselines, achieving superior visual knowledge ablation while robustly preserving general model utility and text-only knowledge.

Given privacy and network load concerns, employing on-device multimodal neural networks (MNNs) for IoT data is a growing trend. However, the high computational demands of MNNs clash with limited on-device resources. MNNs involve input and model redundancies during inference, wasting resources to process redundant input components and run excess model parameters. Model Redundancy Elimination (MRE) reduces redundant parameters but cannot bypass inference for unnecessary input components. Input Redundancy Elimination (IRE) skips inference for redundant input components but cannot reduce computation for the remaining parts. MRE and IRE independently fail to meet the diverse computational needs of multimodal inference. To address these issues, we aim to combine the advantages of MRE and IRE to achieve a more efficient inference. We propose anadaptivemultilevelredundancyelimination framework (AMRE), which supports both IRE and MRE.AMREfirst establishes a collaborative inference mechanism for IRE and MRE. We then propose a multifunctional, lightweight policy model that adaptively controls the inference logic for each instance. Moreover, a three-stage training method is proposed to ensure the performance of collaborative inference inAMRE. We validateAMREin three scenarios, achieving up to 52.91% lower latency, 56.79% lower energy cost, and a slight accuracy gain compared to state-of-the-art baselines.

With the wide application of multi-modal fusion sensing in scenarios such as autonomous driving and human-computer interaction, the privacy security and communication burden caused by massive data uploading need to be solved urgently. Federated Learning (FL) has received significant attention as a privacy-preserving distributed machine learning paradigm. Recent Multi-Modal Federated Learning (MMFL) focuses on addressing modal heterogeneity to enhance accuracy and speed up convergence. However, it overlooks the huge communication overhead in updating complex multi-modal network models, especially in edge environments with limited bandwidth. At the same time, the state-of-the-art communication-efficient FL methods are not customized to the MMFL characteristics. In this paper, we propose the Adaptive Quantization framework for Multi-modal Federated Learning (AQMFL). AQMFL implements decision-level multi-modal fusion locally by using parallel training and model ensemble, supporting its adaptation to modal heterogeneity and flexible deployment. AQMFL can adaptively allocate the number of quantization levels of gradient according to the modal contribution and the heterogeneous communication ability of nodes, which speeds up the system convergence and achieves a better balance between accuracy and communication efficiency. Compared with the classical baselines, AQMFL can reduce the total communication overhead by up to 50.47% and the total training time by up to 52.11% while maintaining the accuracy.

The ubiquitous Internet-of-Things (IoT) devices generate vast amounts of multimodal data, and the deep multimodal fusion network (DMFN) is a promising technology for processing multimodal data. Deploying DMFNs locally on embedded IoT devices is a profitable way to provide privacy-preserving and robust sensing services. However, the current compression methods suffer from the following limitations: First, they are designed based on unimodal networks or specific model structures. Hence, it is hard to extend these methods to diverse DMFNs; Second, existing works never relate their efforts to disparate computational demands of multimodal data and modalities. Easy samples and redundant modalities consume the same computational resources as powerful modalities and complex samples. We propose anAdaptiveCompressionFramework (ACF) for DMFNs to address those challenges. It enables input-dependent runtime compression locally on resource-constrained embedded devices. Specifically, we propose an offline model transformation module to upgrade the static network with two kinds of dynamic components to support online structural adjustment. Then we design a lightweight policy network to generate multi-granularity and data-dependent compression strategies for different model parts. Finally, we evaluate ACF on four DMFNs across three embedded platforms. Compared with the best results of the existing schemes, ACF obtains up to 2.61× latency reduction and 2.30× energy consumption reduction, with up to 3.57% accuracy improvement.

Cloud service providers improve resource utilization by co-locating latency-critical (LC) workloads with best-effort batch (BE) jobs in datacenters. However, they usually treat an LC workload as a whole when allocating resources to BE jobs and neglect the different features of components of an LC workload. This kind of coarse-grained co-location method leaves a significant room for improvement in resource utilization.