Papers

The rapid expansion of distributed rooftop photovoltaic (PV) systems introduces increasing uncertainty in distribution grid planning, hosting capacity assessment, and voltage regulation. Reliable estimation of rooftop PV deployment from satellite imagery is therefore essential for accurate modeling of distributed generation at feeder and service-territory scales. However, conventional computer vision approaches rely on fixed learned representations and globally averaged visual correlations. This makes them sensitive to geographic distribution shifts caused by differences in roof materials, urban morphology, and imaging conditions across regions. To address these challenges, this paper proposes Solar Retrieval-Augmented Generation (Solar-RAG), a context-grounded framework for photovoltaic assessment that integrates similarity-based image retrieval with multimodal vision-language reasoning. Instead of producing predictions solely from internal model parameters, the proposed approach retrieves visually similar rooftop scenes with verified annotations and performs comparative reasoning against these examples during inference. This retrieval-guided mechanism provides geographically contextualized references that improve robustness under heterogeneous urban environments without requiring model retraining. The method outperform both conventional deep vision models and standalone vision-language models. Furthermore, feeder-level case studies show that improved PV inventory estimation reduces errors in voltage deviation analysis and hosting capacity assessment. The results demonstrate that the proposed method provides a scalable and geographically robust approach for monitoring distributed PV deployment. This enables more reliable integration of remote sensing data into distribution grid planning and distributed energy resource management.

Change detection of high-resolution remote sensing images is an important task in earth observation and was extensively investigated. Recently, deep learning has shown to be very successful in plenty of remote sensing tasks. The current deep learning-based change detection method is mainly based on conventional long short-term memory (Conv-LSTM), which does not have spatial characteristics. Since change detection is a process with both spatiality and temporality, it is necessary to propose an end-to-end spatiotemporal network. To achieve this, Conv-LSTM, an extension of the Conv-LSTM structure, is introduced. Since it shares similar spatial characteristics with the convolutional layer, L-UNet, which substitutes partial convolution layers of UNet-to-Conv-LSTM and Atrous L-UNet (AL-UNet), which further using Atrous structure to multiscale spatial information is proposed. Experiments on two data sets are conducted and the proposed methods show the advantages both in quantity and quality when compared with some other methods.

In this paper, we present new results on finite- and fixed-time convergence for dynamical systems using LaSalle-like invariance principles. In particular, we provide first and second-order non-smooth Lyapunov-like results for finite- and fixed-time convergence, thereby relaxing the requirement of existence a differentiable, positive definite Lyapunov function. Based on these findings, we show that a dynamical system whose equilibrium point is globally asymptotically stable can be modified through scaling so that the resulting dynamical system has a fixed-time stable equilibrium point. The results in this paper expand our understanding of various convergence rates and strengthen the hypothesis that all the convergence rates are interconnected through a suitable transformation.

Given the popularity of 360{\deg} images on social media platforms, 360{\deg} image compression becomes a critical technology for media storage and transmission. Conventional 360{\deg} image compression pipeline projects the spherical image into a single 2D plane, leading to issues of oversampling and distortion. In this paper, we propose a novel viewport-based neural compression pipeline for 360{\deg} images. By replacing the image projection in conventional 360{\deg} image compression pipelines with viewport extraction and efficiently compressing multiple viewports, the proposed pipeline minimizes the inherent oversampling and distortion issues. However, viewport extraction impedes information sharing between multiple viewports during compression, causing the loss of global information about the spherical image. To tackle this global information loss, we design a neural viewport codec to capture global prior information across multiple viewports and maximally compress the viewport data. The viewport codec is empowered by a transformer-based ViewPort ConText (VPCT) module that can be integrated with canonical learning-based 2D image compression structures. We compare the proposed pipeline with existing 360{\deg} image compression models and conventional 360{\deg} image compression pipelines building on learning-based 2D image codecs and standard hand-crafted codecs. Results show that our pipeline saves an average of $14.01\%$ bit consumption compared to the best-performing 360{\deg} image compression methods without compromising quality. The proposed VPCT-based codec also outperforms existing 2D image codecs in the viewport-based neural compression pipeline. Our code can be found at: https://github.com/Jingwei-Liao/VPCT.

This paper investigates the problem of pose synchronization for multiple rigid body systems evolving on the matrix Lie group $\SE(3)$. We propose a distributed hybrid feedback control scheme with global asymptotic stability guarantees using relative pose and group velocity measurements. The key idea consists of constructing a new potential function on $\SE(3) \times \mathbb{R}$ with a generalized non-diagonal weighting matrix, and a set of auxiliary scalar variables with continuous-discrete hybrid dynamics. Based on the new potential function and the auxiliary scalar variables, a geometric distributed hybrid feedback designed directly on $\SE(3)$ is proposed to achieve global pose synchronization. Numerical simulation results are presented to illustrate the performance of the proposed distributed hybrid control scheme.

Autonomous mobile robot fleets must coordinate task allocation and charging under limited shared resources, yet most battery aware planning methods address only a single robot. This paper extends degradation cost aware task planning to a multi robot setting by jointly optimizing task assignment, service sequencing, optional charging decisions, charging mode selection, and charger access while balancing degradation across the fleet. The formulation relies on reduced form degradation proxies grounded in the empirical battery aging literature, capturing both charging mode dependent wear and idle state of charge dependent aging; the bilinear idle aging term is linearized through a disaggregated piecewise McCormick formulation. Tight big M values derived from instance data strengthen the LP relaxation. To manage scalability, we propose a hierarchical matheuristic in which a fleet level master problem coordinates assignments, routes, and charger usage, while robot level subproblems whose integer part decomposes into trivially small independent partition selection problems compute route conditioned degradation schedules. Systematic experiments compare the proposed method against three baselines: a rule based nearest available dispatcher, an energy aware formulation that enforces battery feasibility without modeling degradation, and a charger unaware formulation that accounts for degradation but ignores shared charger capacity limits.

This paper presents the Interspeech 2026 Audio Encoder Capability Challenge, a benchmark specifically designed to evaluate and advance the performance of pre-trained audio encoders as front-end modules for Large Audio Language Models (LALMs). While LALMs have shown remarkable understanding of complex acoustic scenes, their performance depends on the semantic richness of the underlying audio encoder representations. This challenge addresses the integration gap by providing a unified generative evaluation framework, XARES-LLM, which assesses submitted encoders across a diverse suite of downstream classification and generation tasks. By decoupling encoder development from LLM fine-tuning, the challenge establishes a standardized protocol for general-purpose audio representations that can effectively be used for the next generation of multimodal language models.

This work proposes a novel immersive communication framework that leverages brain-computer interface (BCI) to acquire brain signals for inferring user-centric states (e.g., intention and perception-related discomfort), thereby enabling more personalized and robust immersive adaptation under strong individual variability. Specifically, we develop a personalized federated learning (PFL) model to analyze and process the collected brain signals, which not only accommodates neurodiverse brain-signal data but also prevents the leakage of sensitive brain-signal information. To address the energy bottleneck of continual on-device learning and inference on energy-limited immersive terminals (e.g., head-mounted display), we further embed spiking neural networks (SNNs) into the PFL. By exploiting sparse, event-driven spike computation, the SNN-enabled PFL reduces the computation and energy cost of training and inference while maintaining competitive personalization performance. Experiments on real brain-signal dataset demonstrate that our method achieves the best overall identification accuracy while reducing inference energy by 6.46$\times$ compared with conventional artificial neural network-based personalized baselines.

In waveguide quantum electrodynamics (QED) systems, a giant cavity can be engineered to interact with quantum fields by multiple distant coupling points so that its non-Markovian dynamics are quite different from traditional quantum optical cavity systems. Towards feedback control this system, this paper designs an optimal filter for the giant cavity systems to estimate its state evolution under continuous quantum measurements. Firstly, the Langevin equation in the Heisenberg picture are derived, which is a linear continuous-time system with both states and inputs delays resulting from the unconventional distant couplings. Compared to existing modeling approaches, this formulation effectively preserves the nonlocal coupling and multiple delay dynamic characteristics inherent in the original system. In particular, the presence of coupling and propagation delays leads to noncommutativity among the system operators at different times, which prevents the direct application of existing quantum filtering methods. To address this issue, an optimal filter is designed, in which the delayed-state covariance matrices are computed. By iteratively evaluating the delayed-state covariance over successive time intervals, the resulting optimal filter can be implemented in an interval-wise backward recursion algorithm. Finally, numerical simulations are conducted to evaluate the tracking performance of the proposed optimal filter for the giant cavity. By comparing between the evolutions of Wigner functions of coherent and cat states and the filter, the effectiveness of the optimal filter is validated.

Conversational automatic speech recognition remains challenging due to overlapping speech, far-field noise, and varying speaker counts. While recent LLM-based systems perform well on single-speaker benchmarks, their robustness in multi-speaker settings is unclear. We systematically compare LLM-based and modular pipeline approaches along four axes: overlap robustness, semantic fidelity, speaker count, and single- versus multi-channel input. To capture meaning-altering errors that conventional metrics miss, we introduce tcpSemER, which extends tcpWER by replacing Levenshtein distance with embedding-based semantic similarity. We further decompose tcpWER into overlapping and non-overlapping components for finer-grained analysis. Experiments across three datasets show that LLM-based systems are competitive in two-speaker settings but degrade as speaker count and overlap increase, whereas modular pipelines remain more robust.

We develop an optimization-free framework for safe stabilization of single-input control-affine nonlinear systems with a given control Lyapunov function (CLF) and a given control barrier function (CBF), where the desired equilibrium lies in the interior of the safe set. An explicit compatibility condition is derived that is necessary and sufficient for the pointwise simultaneous satisfaction of the CLF and CBF inequalities. When this condition holds, two closed-form continuous state-feedback laws are constructed from the Lie-derivative data of the CLF and CBF via standard universal stabilizer formulas, yielding asymptotic stabilization of the origin and forward invariance of the interior of the safe set, without online quadratic programming. The two laws belong to broader families parametrized by a free nondecreasing function, providing additional design flexibility. When the compatibility condition fails, a safety-prioritizing modification preserves forward invariance and drives the state toward the safe-set boundary until a compatible region is reached, whereupon continuity at the origin and asymptotic stabilization are recovered. The framework produces families of explicit constructive alternatives to CLF-CBF quadratic programming for scalar-input nonlinear systems.

This paper studies quantum-encrypted explicit MPC for constrained discrete-time linear systems in a cloud-based architecture. A finite-horizon quadratic MPC problem is solved offline to obtain a piecewise-affine controller. Shared quantum keys generated from Bell pairs and protected by quantum key distribution are used to encrypt the online control evaluation between the sensor and actuator. Based on this architecture, we develop a lightweight encrypted explicit MPC protocol, prove exact recovery of the plaintext control action, and characterize its computational efficiency. Numerical results demonstrate lower online complexity than classical encrypted MPC, while security is discussed in terms of confidentiality of plant data and control inputs.

Talking-head videos constitute a predominant content type in real-time communication, yet publicly available datasets for video processing research in this domain remain scarce and limited in signal fidelity. In this paper, we open-source a camera-native dataset of 838 talking-head recordings (approximately 210 minutes), each 15s in duration, captured from 799 participants across 443 camera-code categories in their natural environments. All recordings are stored using the FFV1 lossless codec, preserving the camera-native signal—uncompressed (24.7%) or MJPEG-encoded (75.3%)—without additional lossy processing. Each recording is annotated with a Mean Opinion Score (MOS) and ten perceptual quality tokens that jointly explain 64.4% of the MOS variance. From this corpus, we curate a stratified benchmarking subset of 120 clips in three content conditions: original, background blur, and background replacement. Codec efficiency evaluation across four datasets and four codecs, namely H.264, H.265, H.266, and AV1, yields VMAF BD-rate savings up to $-71.3%$ (H.266) relative to H.264, with significant encoder$\times$dataset ($\eta_p^2 = .112$) and encoder$\times$content condition ($\eta_p^2 = .149$) interactions, demonstrating that both content type and background processing affect compression efficiency. A preliminary super-resolution evaluation with four SR models confirms that the dataset significantly affects absolute performance while preserving model rankings, demonstrating applicability beyond codec benchmarking. The dataset offers 5$\times$ the scale of the largest prior talking-head webcam dataset (838 vs. 160 clips) and preserves the camera-native signal without additional lossy compression, establishing a resource for benchmarking video compression, super-resolution, quality assessment, and enhancement models in real-time communication.

Despite their output being ultimately consumed by human viewers, 3D Gaussian Splatting (3DGS) methods often rely on ad-hoc combinations of pixel-level losses, resulting in blurry renderings. To address this, we systematically explore perceptual optimization strategies for 3DGS by searching over a diverse set of distortion losses. We conduct the first-of-its-kind large-scale human subjective study on 3DGS, involving 39,320 pairwise ratings across several datasets and 3DGS frameworks. A regularized version of Wasserstein Distortion, which we call WD-R, emerges as the clear winner, excelling at recovering fine textures without incurring a higher splat count. WD-R is preferred by raters more than $2.3\times$ over the original 3DGS loss, and $1.5\times$ over the current best method Perceptual-GS. WD-R also consistently achieves state-of-the-art LPIPS, DISTS, and FID scores across various datasets, and generalizes across recent frameworks, such as Mip-Splatting and Scaffold-GS, where replacing the original loss with WD-R consistently enhances perceptual quality within a similar resource budget (number of splats for Mip-Splatting, model size for Scaffold-GS), and leads to reconstructions being preferred by human raters $1.8\times$ and $3.6\times$, respectively. We also find that this carries over to the task of 3DGS scene compression, with $\approx 50\%$ bitrate savings for comparable perceptual metric performance.

Clinical MRI frequently acquires anisotropic volumes with high in-plane resolution and low through-plane resolution to reduce acquisition time. Multiple orientations are therefore acquired to provide complementary anatomical information. Conventional integration of these views relies on registration followed by interpolation, which can degrade fine structural details. Recent deep learning-based super-resolution (SR) approaches have demonstrated strong performance in enhancing single-view images. However, their clinical reliability is often limited by the need for large-scale training datasets, resulting in increased dependence on cohort-level priors. Self-supervised strategies offer an alternative by learning directly from the target scans. Prior work either neglects the existence of multi-view information or assumes that in-plane information can supervise through-plane reconstruction under the assumption of pre-alignment between images. However, this assumption is rarely satisfied in clinical settings. In this work, we introduce Single-Subject Implicit Multi-View Super-Resolution for MRI (SIMS-MRI), a framework that operates solely on anisotropic multi-view scans from a single patient without requiring pre- or post-processing. Our method combines a multi-resolution hash-encoded implicit representation with learned inter-view alignment to generate a spatially consistent isotropic reconstruction. We validate the SIMS-MRI pipeline on both simulated brain and clinical prostate MRI datasets. Code will be made publicly available for reproducibility: https://github.com/abhshkt/SIMS-MRI

With the increasing deployment of automated and agentic systems, ensuring the adversarial robustness of automatic speech recognition (ASR) models has become highly relevant. We observe that changing the precision of an ASR model during inference reduces the likelihood of adversarial attacks to succeed. We take advantage of this fact to make models more robust simply by randomly sampling the precision during prediction. Moreover, this insight can be turned into an adversarial example detection strategy by implementing a simple Gaussian classifier that thresholds the differences between outputs of models run with different precision. To further enhance security boundaries, we combine the approach with an existing uncertainty-based defense mechanism, which forces adaptive adversaries to introduce highly perceptible noise to bypass detection. An experimental analysis across various ASR models, languages, and attack types demonstrates a significant increase in adversarial robustness, competitive detection capabilities, and resistance to adaptive threats.

First-order Ambisonics (FOA) is a standard spatial audio format based on spherical harmonic decomposition. Its zeroth- and first-order components capture the sound pressure and particle velocity, respectively. Recently, physics-informed neural networks have been applied to the spatial interpolation of FOA signals, regularizing the network outputs based on soft penalty terms derived from physical principles, e.g., the linearized momentum equation. In this paper, we reformulate the task so that the predicted FOA signal automatically satisfies the linearized momentum equation. Our network approximates a scalar function called velocity potential, rather than the FOA signal itself. Then, the FOA signal can be readily recovered through the partial derivatives of the velocity potential with respect to the network inputs (i.e., time and microphone position) according to physics of sound propagation. By deriving the four channels of FOA from the single-channel velocity potential, the reconstructed signal follows the physical principle at any time and position by construction. Experimental results on room impulse response reconstruction confirm the effectiveness of the proposed framework.

Agrivoltaic systems–photovoltaic (PV) panels installed above agricultural land–have emerged as a promising dual-use solution to address competing land demands for food and energy production. In this paper, we propose a model predictive control (MPC) approach to dual-axis agrivoltaic panel tracking control that dynamically adjusts panel positions in real time to maximize power production and crop yield given solar irradiance and ambient temperature measurements. We apply convex relaxations and shading factor approximations to reformulate the MPC optimization problem as a convex second-order cone program that determines the PV panel position adjustments away from the sun-tracking trajectory. Through case studies, we demonstrate our approach, exploring the Pareto front between i) an approach that maximizes power production without considering crop needs and ii) crop yield with no agrivoltaics. We also conduct a case study exploring the impact of forecast error on MPC performance. We find that dynamically adjusting agrivoltaic panel position helps us actively manage the trade-offs between power production and crop yield, and that active panel control enables the agrivoltaic system to achieve land equivalent ratio values of up to 1.897.

The adversarial subproblem in two-stage adaptive robust optimization (ARO), which identifies the worst-case uncertainty realization, is a major computational bottleneck. This difficulty is exacerbated when the recourse value function is non-concave and the uncertainty set shifts across applications. Existing approaches typically exploit specific structural assumptions on the value function or the uncertainty set geometry to reformulate this subproblem, but degrade when these assumptions are violated or the geometry changes at deployment. To address this challenge, we propose L2O-CCG, a bi-level framework that enables the integration of structure-aware adversarial solvers within the constraint-and-column generation (CCG) algorithm. As one instantiation, we develop a generalizable adversarial learning method, which replaces solver-based adversarial search with a learned proximal gradient optimizer that can generalize across uncertainty set geometries without retraining. Here, an inner-level neural network approximates the recourse value function from offline data, while an outer-level pre-trained mapping generates iteration-dependent step sizes for a proximal gradient scheme. We also establish out-of-distribution convergence bounds under uncertainty set parameter shifts, showing how the trajectory deviation of the learned optimizer is bounded by the uncertainty set shift. We illustrate performance of the L2O-CCG method on a building HVAC management task.

Affective computing aims to understand and model human emotions for computational systems. Within this field, speech emotion recognition (SER) focuses on predicting emotions conveyed through speech. While early SER systems relied on limited datasets and traditional machine learning models, recent deep learning approaches demand largescale, naturalistic emotional corpora. To address this need, we introduce the MSP-Conversation corpus: a dataset of more than 70 hours of conversational audio with time-continuous emotional annotations and detailed speaker diarizations. The time-continuous annotations capture the dynamic and contextdependent nature of emotional expression. The annotations in the corpus include fine-grained temporal traces of valence, arousal, and dominance. The audio data is sourced from publicly available podcasts and overlaps with a subset of the isolated speaking turns in the MSP-Podcast corpus to facilitate direct comparisons between annotation methods (i.e., in-context versus out-of-context annotations). The paper outlines the development of the corpus, annotation methodology, analyses of the annotations, and baseline SER experiments, establishing the MSP-Conversation corpus as a valuable resource for advancing research in dynamic SER in naturalistic settings.

Mapping is essential in robotics and autonomous systems because it provides the spatial foundation for path planning. Efficient mapping enables planning algorithms to generate reliable paths while ensuring safety and adapting in real time to complex environments. Fixed-resolution mapping methods often produce overly conservative obstacle representations that lead to suboptimal paths or planning failures in cluttered scenes. To address this issue, we introduce Parallel OctoMapping (POMP), an efficient OctoMap-based mapping technique that maximizes available free space and supports multi-threaded computation. To the best of our knowledge, POMP is the first method that, at a fixed occupancy-grid resolution, refines the representation of free space while preserving map fidelity and compatibility with existing search-based planners. It can therefore be integrated into existing planning pipelines, yielding higher pathfinding success rates and shorter path lengths, especially in cluttered environments, while substantially improving computational efficiency.

Future wireless networks are expected to support increasingly high data rates and user densities, motivating advanced multi-antenna architectures capable of adapting to dynamic propagation environments. Movable antenna (MA) arrays have recently emerged as an extension of massive MIMO, enabling physical repositioning of antenna elements to better exploit spatial diversity and mitigate inter-user interference. While prior studies report promising gains under idealized assumptions, their performance under realistic wideband multi-user operation remains insufficiently understood. This paper presents a comprehensive evaluation of MA-enabled systems in practical uplink and downlink scenarios. A wideband OFDM system model is developed, and novel closed-form sum-rate expressions are derived for both uplink and downlink under linear and nonlinear processing. Hardware impairments are incorporated via an EVM-based model, from which a distortion-aware UL/DL duality is established and the resulting high-SNR sum rate ceiling is analytically characterized. In addition, the interactions between antenna position optimization, receiver processing, and user loading are examined, and performance is evaluated under both time-division duplexing (TDD) and frequency-division duplexing (FDD). The results show that movable antennas can provide noticeable gains in low-impairment regimes with strong multi-user interference, but these benefits are highly scenario-dependent and diminish under hardware-impairment-limited conditions or in rich-scattering environments. These findings highlight the importance of carefully assessing deployment conditions when considering antenna mobility as an alternative to conventional fixed array configurations.

We present a compressive beamforming method for coherent plane-wave compounding (CPWC) ultrasound imaging based on a far-field decomposition of the received radiofrequency (RF) data into virtual plane waves. This decomposition recasts the imaging operation entirely in the spatial frequency domain ($k$-space), allowing direct and flexible control over $k$-space sampling distributions based on the principle of coarrays. We present vernier-type sampling strategies designed to optimize the tradeoff between image contrast and resolution with minimum redundancy, including strategies that favor dense low-frequency sampling for high contrast, shifted schemes that extend the frequency support for improved resolution, and confocal or hybrid compounding schemes that approximate the spatial-frequency transfer function of conventional DAS beamforming. Our method, called KK beamforming, is validated with a calibration phantom and in-vivo human tissue data, demonstrating compression factors of an order of magnitude while maintaining image qualities comparable to conventional DAS. We further demonstrate that KK beamforming yields improvements in computational speed owing to its reduced memory footprint and more efficient cache utilization of the compressed data and associated look-up tables.

The growing complexity of modern control tasks calls for controllers that can react online as objectives and disturbances change, while preserving closed-loop stability. Recent approaches for improving the performance of nonlinear systems while preserving closed-loop stability rely on time-invariant recurrent neural-network controllers, but offer no principled way to update the controller during operation. Most importantly, switching from one stabilizing policy to another can itself destabilize the closed-loop. We address this problem by introducing a stability-preserving update mechanism for nonlinear, neural-network-based controllers. Each controller is modeled as a causal operator with bounded $\ell_p$-gain, and we derive gain-based conditions under which the controller may be updated online. These conditions yield two practical update schemes, time-scheduled and state-triggered, that guarantee the closed-loop remains $\ell_p$-stable after any number of updates. Our analysis further shows that stability is decoupled from controller optimality, allowing approximate or early-stopped controller synthesis. We demonstrate the approach on nonlinear systems with time-varying objectives and disturbances, and show consistent performance improvements over static and naive online baselines while guaranteeing stability.

This paper develops a method to construct robust positively invariant (RPI) tube sets from finite noisy input-state data of an unknown linear time-invariant (LTI) system, yielding tubes that can be directly embedded in tube-based robust data-driven predictive control. Data-consistency uncertainty sets are constructed under process/measurement noise with polytopic/ellipsoidal bounds. In the measurement-noise case, we provide a deterministic and data-consistent procedure to certify the induced residual bound from data. Based on these sets, a robustly stabilizing state-feedback gain is certified via a common quadratic contraction, which in turn enables constructive polyhedral/ellipsoidal RPI tube computation. Numerical examples quantify the conservatism induced by noisy data and the employed certification step.

This paper studies the problem of finite-time convergence to a prescribed safe set for nonlinear systems whose initial states violate the safety constraints. Existing Control Lyapunov-Barrier Functions (CLBFs) can enforce recovery to the safe set but may suffer from the issue of chattering and they do not explicitly consider control bounds. To address these limitations, we propose a new Control Barrier Function (CBF) formulation that guarantees finite-time convergence to the safe set while ensuring feasibility under control constraints. Specifically, we strengthen the initially violated safety constraint by introducing a parameter which enables the exploitation of the asymptotic property of a CBF to converge to the safe set in finite time. Furthermore, the conditions for the existence of such a CBF under control bounds to achieve finite-time convergence are derived via reachability analysis and constraint comparison, providing a systematic approach for parameter design. A case study on 2D obstacle avoidance is presented to demonstrate the effectiveness and advantages of the proposed method.

We present mmFHE, the first system that executes the entire cloud-side mmWave sensing pipeline including the DSP and ML inference under fully homomorphic encryption (FHE). mmFHE encrypts range profiles on an edge device after lightweight plaintext preprocessing and executes the entire mmWave signal-processing and ML inference pipeline homomorphically on a semi-honest cloud that operates exclusively on ciphertexts. At the core of mmFHE is a library of seven composable, data-oblivious FHE kernels that replace standard DSP routines with fixed arithmetic circuits for different application-specific pipelines. We demonstrate this approach on two representative tasks: vital-sign monitoring and gesture recognition. We formally prove two cryptographic guarantees for any pipeline assembled from this library: input privacy and data obliviousness. These guarantees effectively neutralize various supervised and unsupervised privacy attacks on raw data, including re-identification and data-dependent privacy leakage. Evaluation on three public radar datasets shows that encryption introduces negligible error versus the plaintext pipeline, with 84.5% gesture accuracy (vs. 84.7%). End-to-end cloud GPU latency is 1.21 s per 10 s vital-sign window and 5.76 s per 3 s gesture window. These results establish the initial feasibility of end-to-end mmWave sensing under FHE on commodity hardware.

Gastrointestinal (GI) tract image analysis plays a crucial role in medical diagnosis. This research addresses the challenge of accurately classifying and segmenting GI images for real-time applications, where traditional methods often struggle due to the diversity and complexity of abnormalities. The high computational demands of this domain require efficient and adaptable solutions. This PhD thesis presents a multifaceted approach to GI image analysis. Initially, texture-based feature extraction and classification methods were explored, achieving high processing speed (over 4000 FPS) and strong performance (F1-score: 0.76, Accuracy: 0.98) on the Kvasir V2 dataset. The study then transitions to deep learning, where an optimized model combined with data bagging techniques improved performance, reaching an accuracy of 0.92 and an F1-score of 0.60 on the HyperKvasir dataset, and an F1-score of 0.88 on Kvasir V2. To support real-time detection, a streamlined neural network integrating texture and local binary patterns was developed. By addressing inter-class similarity and intra-class variation through a learned threshold, the system achieved 41 FPS with high accuracy (0.99) and an F1-score of 0.91 on HyperKvasir. Additionally, two segmentation tools are proposed to enhance usability, leveraging Depth-Wise Separable Convolution and neural network ensembles for improved detection, particularly in low-FPS scenarios. Overall, this research introduces novel and adaptable methodologies, progressing from traditional texture-based techniques to deep learning and ensemble approaches, providing a comprehensive framework for advancing GI image analysis.

Video-based gait analysis has become a promising approach for assessing motor impairment in children with cerebral palsy (CP). However, existing methods usually rely on either pose sequences or handcrafted gait features alone, making it difficult to simultaneously capture spatiotemporal motion patterns and clinically meaningful biomechanical information. To address this gap, we propose a multimodal fusion framework that integrates skeleton dynamics with contribution-guided clinically meaningful gait features. First, Grad-CAM analysis on a pre-trained ST-GCN backbone identified the most discriminative body keypoints, providing an interpretable basis for subsequent gait feature extraction. We then build a dual-stream architecture, with one stream modeling skeleton dynamics using ST-GCN and the other encoding gait geatures derived from the identified keypoints. By fusing the two streams through feature cross-attention improved four-level CP motor severity classification to 70.86%, outperforming the baseline by 5.6 percentage points. Overall, this work suggests that integrating skeleton dynamics with clinically meaningful gait descriptors can improve both prediction performance and biomechanical interpretability for video-based CP severity assessment.

We introduce TiCo, a time-controllable spoken dialogue model (SDM) that follows time-constrained instructions (e.g., "Please generate a response lasting about 15 seconds") and generates spoken responses with controllable duration. This capability is valuable for real-world spoken language systems such as voice assistants and interactive agents, where controlling response duration can improve interaction quality. However, despite their strong ability to generate natural spoken responses, existing models lack time awareness and struggle to follow duration-related instructions. To systematically evaluate this, we introduce TiCo-Bench, the first benchmark for time-controllable instruction following in SDMs, on which existing open-source and commercial models frequently fail to satisfy explicit time constraints. TiCo addresses this limitation by enabling an SDM to estimate elapsed speaking time during generation through Spoken Time Markers (STM) (e.g., <10.6 seconds>). These markers help the model maintain awareness of time and adjust the remaining content to meet the target duration. TiCo is post-trained efficiently without question-answer paired data, relying on self-generation and reinforcement learning with verifiable reward. Experimental results show that TiCo reduces duration error by 2.7x over its backbone and 1.6x over the strongest baseline, while preserving response quality.