Large Language Model (LLM) agent systems have advanced rapidly, driven by their strong generalization in zero-shot settings. To further enhance reasoning and accuracy on complex tasks, Multi-Agent Debate (MAD) has emerged as a promising framework that engages multiple LLM agents in structured debates to encourage diverse reasoning. However, triggering MAD for every query is inefficient, as it incurs substantial computational (token) cost and may even degrade accuracy by overturning correct answers from single-agent. To address these limitations, we propose intelligent Multi-Agent Debate (iMAD), a token-efficient framework that selectively triggers MAD only when it is likely to be beneficial (i.e., correcting an initially wrong answer). To achieve this goal, iMAD learns generalizable model behaviors to make accurate debate decisions. Specifically, iMAD first prompts a single agent to produce a structured self-critique response, from which we extract 41 interpretable linguistic and semantic features capturing hesitation cues. Then, iMAD uses a lightweight debate decision classifier, trained using our proposed FocusCal loss without test-dataset-specific tuning, to make robust zero-shot debate decisions. Through extensive experiments using six (visual) question answering datasets against five competitive baselines, we show that iMAD significantly reduces token usage (by up to 92%) while also improving final answer accuracy (by up to 13.5%).

High-resolution Vision-Language Models (VLMs) are widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate an excessive number of visual tokens due to the need to encode multiple partitions of a high-resolution image input. Processing such a large number of visual tokens poses significant computational challenges, particularly for resource-constrained commodity GPUs. To address this challenge, we propose High-Resolution Early Dropping (HiRED), a plug-and-play token-dropping method designed to operate within a fixed token budget. HiRED leverages the attention of CLS token in the vision transformer (ViT) to assess the visual content of the image partitions and allocate an optimal token budget for each partition accordingly. The most informative visual tokens from each partition within the allocated budget are then selected and passed to the subsequent Large Language Model (LLM). We showed that HiRED achieves superior accuracy and performance, compared to existing token-dropping methods. Empirically, HiRED-20% (i.e., a 20% token budget) on LLaVA-Next-7B achieves a 4.7x increase in token generation throughput, reduces response latency by 78%, and saves 14% of GPU memory for single inference on an NVIDIA TESLA P40 (24 GB). For larger batch sizes (e.g., 4), HiRED-20% prevents out-of-memory errors by cutting memory usage by 30%, while preserving throughput and latency benefits.

Split Learning (SL) is an emerging privacy-preserving machine learning technique that enables resource constrained edge devices to participate in model training by partitioning a model into client-side and server-side sub-models. While SL reduces computational overhead on edge devices, it encounters significant challenges in heterogeneous environments where devices vary in computing resources, communication capabilities, environmental conditions, and privacy requirements. Although recent studies have explored heterogeneous SL frameworks that optimize split points for devices with varying resource constraints, they often neglect personalized privacy requirements and local model customization under varying environmental conditions. To address these limitations, we propose P3SL, a Personalized Privacy-Preserving Split Learning framework designed for heterogeneous, resource-constrained edge device systems. The key contributions of this work are twofold. First, we design a personalized sequential split learning pipeline that allows each client to achieve customized privacy protection and maintain personalized local models tailored to their computational resources, environmental conditions, and privacy needs. Second, we adopt a bi-level optimization technique that empowers clients to determine their own optimal personalized split points without sharing private sensitive information (i.e., computational resources, environmental conditions, privacy requirements) with the server. This approach balances energy consumption and privacy leakage risks while maintaining high model accuracy. We implement and evaluate P3SL on a testbed consisting of 7 devices including 4 Jetson Nano P3450 devices, 2 Raspberry Pis, and 1 laptop, using diverse model architectures and datasets under varying environmental conditions. Experimental results demonstrate that P3SL significantly mitigates privacy leakage risks, reduces system energy consumption by up to 59.12%, and consistently retains high accuracy compared to the state-of-the-art heterogeneous SL system.

With the advancement of Artificial Intelligence (AI) towards multiple modalities (language, vision, speech, etc.), multi-modal models have increasingly been used across various applications (e.g., visual question answering or image generation/captioning). Despite the success of AI as a service for multi-modal applications, it relies heavily on clouds, which are constrained by bandwidth, latency, privacy concerns, and unavailability under network or server failures. While on-device AI becomes popular, supporting multiple tasks on edge devices imposes significant resource challenges. To address this, we introduce S2M3, a split-and-share multi-modal architecture for multi-task inference on edge devices. Inspired by the general-purpose nature of multi-modal models, which are composed of multiple modules (encoder, decoder, classifier, etc.), we propose to split multi-modal models at functional-level modules; and then share common modules to reuse them across tasks, thereby reducing resource usage. To address cross-model dependency arising from module sharing, we propose a greedy module-level placement with per-request parallel routing by prioritizing compute-intensive modules. Through experiments on a testbed consisting of 14 multi-modal models across 5 tasks and 10 benchmarks, we demonstrate that S2M3 can reduce memory usage by up to 50% and 62% in single-task and multi-task settings, respectively, without sacrificing accuracy. Furthermore, S2M3 achieves optimal placement in 89 out of 95 instances (93.7%) while reducing inference latency by up to 56.9% on resource-constrained devices, compared to cloud AI.

With the rise of intelligent services on edge devices, the focus of intelligence formation has shifted to the user-side, enabling faster and customized services near the data source, without network overhead or privacy concerns. However, on-device edge intelligence faces some challenges caused by limited data availability and resource constraints in computation and memory. Fortunately, there is more and more intelligence nearby. To effectively harness the potential of widespread edge intelligence, we introduce CollectiveFL, a federated edge intelligence framework that facilitates de-biased, robust edge-to-edge knowledge transfer. By tailoring knowledge sharing for each device based on decision logic similarity, we ensure that edge-side learners specialize in their respective purposes and leverage purpose-specific data. Importantly, to mitigate some possible biases on their own or transferred local data, we delegate knowledge transfer to a set of selected neighboring devices rather than one. By sharing and consolidating collective yet customized intelligence, CollectiveFL establishes collaborative edge-only intelligence without the help of remote servers. Our extensive experimental results on four different model architectures using 13 public datasets have demonstrated that CollectiveFL enhances local learning in 55 out of 63 cases (87.3%) while improving the accuracy of individual tasks by up to 14.1%.

As wireless-equipped devices are widely deployed, fingerprint-based indoor localization becomes popular due to its simple yet precise feature. A key challenge is constructing an accurate map of signals with their corresponding coordinates. However, because the structural layout of each location uniquely affects signal propagation from distinct access points (APs), fingerprint maps cannot be transferred to other locations. This leads to localization failure in unexplored areas. In this paper, we propose CollageMap, an obstacle-aware fingerprint map constructor embracing generic signal features and AP-oriented unique features. We tackle the problem of fingerprint construction as a compound of two complementary maps: 1) obstacle-independent universal map reflecting intrinsic propagation patterns; and 2) obstacle-dependent adaptation map representing the extrinsic effect of obstacles. We construct a universal model that learns existing fingerprints in various training locations so that it can be generally used at any other place. On top of the universal map, another deep neural network (DNN) learns the real signal deviations between the universal map and the ground-truth map and generates the compensation as the adaptation map for obstructed environments. Using real-world received signal strength indicator (RSSI) testbeds across various wireless radios, we have validated CollageMap provides outstanding signal pattern estimation even in the presence of obstacles, achieving improvements in localization accuracy of up to 30.36%, 17.95%, and 16.97% using Wi-Fi, ZigBee, and BLE, respectively, via adaptation. CollageMap effectively keeps the performance gap of only 0.42%, 17.43 and 7.10% on average, compared to the ground-truth map obtained from the site survey.

Indoor localization using deep learning has emerged as a promising approach due to its high accuracy in mapping and predicting user locations for complex datasets. However, the inherent complexity of deep learning models often limits their interpretability, creating a gap in user trust and understanding. This paper introduces eXLoc, a novel framework that integrates Explainable AI, Class Activation Mapping (CAM), into deep learning models for indoor localization to enhance model transparency and interpretability. We introduce a new metric called Impact Score to identify significant APs that affect model predictions. This enhances model interpretability and allows a model to identify the influential APs via their impact on localization performance. We have extensively evaluated eXLoc over eight different places from two real-world RSSI datasets. We gained insights into how the model generates predictions, and identified the reasons for the model’s poor performance. These results demonstrate that our approach can be effectively utilized in enabling users to have more trust and understanding of the model in many real-world scenarios.

Deep learning models deployed on edge devices leverage locally collected data to extract intelligence, mitigating privacy concerns associated with external data sharing. Edge federated learning, an on-device learning paradigm, has emerged as a promising solution, allowing edge nodes to train models locally and share only the trained weights, preserving data privacy. However, it also poses critical challenges of network burden and potential model poisoning. We introduce a self-debiasing security framework Breakwater for multi-hop edge federated learning. We incorporate on-device malicious weight discriminator at each participant, enhancing security and robustness of the federated learning process. The framework strategically balances the benefits of participating nodes with timely defenses against potential malicious clients. Based on the discriminator, we further embed a self-debiasing mechanism that can determine whether to retain or discard the weight propagation from its child nodes. Our Breakwater framework identifies and filters out harmful weights, ensuring the integrity of the global model. Our work contributes to the ongoing discourse on federated learning security, presenting a solution that maintains efficiency while robustly defending against model poisoning threats. We demonstrate its efficacy in enhancing the reliability of the multi-hop edge federated learning process with recovery of up to 69 % in accuracy under attack, offering a path toward secure and cooperative distributed learning environments.

As indoor localization becomes a necessity to provide intelligent location-based services for in-building users, fingerprint-based positioning has been widely adopted in numerous Wi-Fi-equipped devices. However, its reliance on extensive offline site surveys limits its application in unexplored environments without prior fingerprint sampling. To address the challenge, we propose GAN-Loc, a novel generative fingerprint-based frame-work. We extract the underlying correlation between location and radio signal features, empowering indoor localization in un/under-explored areas, including unknown data points, newly deployed APs, or unexplored sites. It involves: 1) decomposing into a signal feature map for each AP perspective; 2) learning with a set of points and their associated signal strength data; 3) generating and integrating synthetic radio fingerprints; and 4) employing them into some existing localization algorithms. We evaluated GAN-Loc with extensive real-world RSSI measurements in seven different indoor places. GAN-Loc achieves the localization accuracy of up to 1.21m, 1.29m, and 1.28m for Wi-Fi, Zigbee, and BLE, respectively, compared to the accuracy of 1.47m, 1.58m, and 1.22m using an ideal ground-truth map, which is unachievable without site survey in unknown sites.

In the era of artificial intelligence (AI), deep neural networks (DNNs) become larger using a massive amount of data, and thus, they are trained via cooperative computing devices (e.g., GPUs or servers) based on federated learning. As computation and data generation move to the edge due to privacy, latency, or bandwidth issue, DNN with edge devices has been investigated. However, edge devices are wirelessly connected and mostly incur fragile connectivity. We propose VersatileFL, a novel volatility-resilient deep learning framework under hostile environments. We address short-term and long-term volatility: 1) versatile distributed learning against short-term fluctuation by substituting the missing intermediate values with the past or approximated values and 2) model rearrangement with runtime connectivity diagnosis against long-term variation by adaptively adjusting the partitioned model for the impaired. We have demonstrated that VersatileFL has achieved 62.0% and 31.9% higher performance than hostile learning without a maintenance scheme against the short-term and long-term volatility, respectively.

As intelligence recently moves to the edge to tackle the problems of privacy, scalability, and network bandwidth in the centralized intelligence, it is necessary to construct an efficient yet robust deep learning model viable at edge devices, which are usually volatile in wireless links and device functionality. The intensive computation burden for deep learning at the edge side necessitates some level of parallel processing via acceleration. We proposeEdgePipe, a deep learning framework based on deep neural networks (DNNs) with a mixture of model parallelism and pipeline training for high resource utilization over volatile wireless edge devices. To tackle the volatility problem in wireless links and device functionality, a concept ofsuper neuronis defined to be a group of neurons across adjacent layers, which is the basis of model partitioning at edge devices. The relatively loss-resilient neuron structure prevents the entire forward or backward training paths from being totally broken down due to only some intermittent link or device failure caused by one or few devices. Furthermore, we design a subsequent pipeline training mechanism based on the prior super-neuron-based model partitioning for fast convergence with more training data in a fixed timeline. The experimental results have demonstrated thatEdgePipeoutperforms several counterpart algorithms includingPipeDreamunder the volatile wireless lossy or device malfunctioning environments, while preserving the low interlayer communication overhead.

In the era of edge computing and Artificial Intelligence (AI), securing billions of edge devices within a network against intelligent attacks is crucial. We propose PUFGAN, an innovative machine learning attack-proof security architecture, by embedding a self-adversarial agent within a device fingerprint- based security primitive, public PUF (PPUF) known for its strong fingerprint-driven cryptography. The self-adversarial agent is implemented using Generative Adversarial Networks (GANs). The agent attempts to self-attack the system based on two GAN variants, vanilla GAN and conditional GAN. By turning the attacking quality through generating realistic secret keys used in the PPUF primitive into system vulnerability, the security architecture is able to monitor its internal vulnerability. If the vulnerability level reaches at a specific value, PUFGAN allows the system to restructure its underlying security primitive via feedback to the PPUF hardware, maintaining security entropy at as high a level as possible. We evaluated PUFGAN on three different machine environments: Google Colab, a desktop PC, and a Raspberry Pi 2, using a real-world PPUF dataset. Extensive experiments demonstrated that even a strong device fingerprint security primitive can become vulnerable, necessitating active restructuring of the current primitive, making the system resilient against extreme attacking environments.

In this paper, we address a route reconstruction problem using Unmanned Aerial Vehicles (UAVs) after a large-scale disaster where stationary ad-hoc networks are severely destructed. The main goal of this paper is to improve routing performance in a progressive manner by reconnecting partitioned networks through dispatched UAV relays. Our proposed algorithm uses two types of UAVs: global and local UAVs to collaboratively find the best deployment position in a dynamically changing environment. To obtain terrestrial network connectivity information and extract high-level network topology, we exploit the concept of strongly connected component in graph theory. Based on the understanding from a global point view, global UAVs recommend the most effective deployment positions to local UAVs so that they are deployed as relays in more critically disrupted areas. Simulation-based experiments validate that our distributed route reconstruction algorithm outperforms a counterpart algorithm in terms of steady-state and dynamic routing performance.

We consider the problem of path planning using multiple UAVs as message ferries to deliver delay-sensitive information in a catastrophic disaster scenario. Our main goal is to find the optimal paths of UAVs to maximize the number of nodes that can successfully be serviced within each designated packet deadline. At the same time, we want to reduce total travel time for visiting over a virtual grid topology. We propose a distributed path planning algorithm that determines the next visit grid point based on a weighted sum of travel time and delivery deadline. Together with path planning, we incorporate a task division mechanism that collaboratively distributes the unvisited grid points with other UAVs so that the entire travel time can substantially be reduced. Simulation results demonstrate that our distributed path planning algorithm mixed with task division outperforms all baseline counterpart algorithms in terms of on-time service node rate and total travel time.