Local Model Research Articles

Augmented Multi-Party Computation Against Gradient Leakage in Federated Learning

Multi-Party Computation (MPC) provides an effective cryptographic solution for distributed computing systems so that local models with sensitive information are encrypted before sending to the centralized servers for aggregation. Though direct local knowledge leakages are eliminated in MPC-based algorithms, we observe the server can still obtain the local information indirectly in many scenarios, or even reveal the groundtruth images through methods like Deep Leakage from Gradients (DLG). To eliminate such possibilities and provide stronger protections, we propose an augmented MPC approach by encrypting local models with two rounds of decomposition before transmitting to the server. The proposed solution allows us to remove the constraint that servers must be honest in the general federated learning settings since the true global model is hidden from the servers. Specifically, the augmented MPC algorithm encodes local models into multiple secret shares in the first round, then each share is furthermore split into a public share and a private share. Consequences of such a two-round decomposition are that the augmented algorithm fully inherits the advantages of standard MPC by providing lossless encryption and decryption while simultaneously rendering the global model invisible to the central server. Both theoretical analysis and experimental verification demonstrate that such an augmented solution can provide stronger protections for the security and privacy of the training data, with minimal extra communication and computation costs incurred.

Accuracy assessment of recent global ocean tide models in coastal waters of the European North West Shelf

The accuracy of global ocean tide models is assessed in coastal waters of the European North West Shelf to ascertain where higher resolution local (forecast) models are most needed for geophysical and navigational applications, and which global models are most suitable for providing boundary conditions for regional and local tide models. Five recent global ocean tide models (FES2014b, EOT20, TPXO9-atlas-v5, GOT4.10c, and DTU16) are considered, with the models first compared by interpolating them onto common grids and computing the mean absolute deviation at each grid point. Coastline tide gauge and offshore bottom pressure sensor data were collated from several sources to give a total of 279 observation sites for evaluating model accuracy, including observational values from 137 locations that have not previously been released and have therefore not been assimilated into any of the global models tested. The residual errors between each model’s predicted phasor and the corresponding observed phasor were calculated at each observation location, and quantified using the root mean square (RMS) and median absolute residual (MAR) for the eight tidal constituents M2, S2, N2, O1, K1, K2, P1, and Q1. To avoid RMS values being biased by observation point density, a Voronoi-weighted RMS based on the water area of the Voronoi polygon about each observation location was also developed and used. Four zones were defined based on ocean depth to gauge model performance, and model inaccuracy is again demonstrated in near-shore regions. Seven further zones were defined based on geographical areas, which reveals inhomogeneity among the global models. The smallest overall root sum square (RSS) RMS error across all eight constituents arises with FES2014b, although TPXO9-atlas-v5 has the best performance when using the MAR and Voronoi-weighted RMS metrics. Using only the 137 observation sites that have not been assimilated by any model and therefore provide an independent accuracy assessment, FES2014b exhibits the smallest errors at the coastline, with an RSS RMS of 24.46 cm. All models exhibit larger errors with the 137 independent observation sites than with all 279 observation sites, with an average overall increase in RSS RMS error of 12%, and an increase of 30% for coastline tide gauges, highlighting the need for local model development in these areas.

Modeling ground vibration excited by a high-speed train running along a viaduct railway using a new structure-soil interaction method

Ground vibrations generated by high-speed trains running along railways either at grade or on viaduct bridges, are greatly concerned by the railway industry. Considering the train-track-bridge-abutment-pile-soil interactions involved, modeling the problem in an analytical manner is almost impossible, while conventional numerical methods struggle to accurately simulate the dynamic properties of infinitely large soils. Therefore, this study introduces a new substructure methodology for train-track-bridge-abutment-pile-soil interactions that leverages an ingenious combination of the 2.5-dimensional finite element-boundary element method (2.5D FE-BE) for dynamics of infinitely long periodic structures, and a local 3-dimensional finite element model for pile-soil interaction. This innovative method allows for the consideration of detailed pile-soil coupling using a local finite element model, all while retaining the ability to account for the infinite size of the soil structure using the 2.5-dimensional finite element-boundary elements. After a comparison between model prediction and field measurement, the model is employed to explore the environmental vibration characteristics of on-viaduct high-speed railways and analyze the vibration reduction effects of various measures, including rail fasteners, track slab pads, and bridge bearings.

Boosting Weakly Supervised Object Localization and Segmentation With Domain Adaption.

Weakly supervised object localization (WSOL), adopting only image-level annotations to learn the pixel-level localization model, can release human resources in the annotation process. Most one-stage WSOL methods learn the localization model with multi-instance learning, making them only activate discriminative object parts rather than the whole object. In our work, we attribute this problem to the domain shift between the training and test process of WSOL and provide a novel perspective that views WSOL as a domain adaption (DA) task. Under this perspective, a DA-WSOL pipeline is elaborated to better assist WSOL with DA approaches by considering the specificities for the adaption of WSOL. Our DA-WSOL pipeline can discern the source-related and the Universum samples from other target samples based on a proposed target sampling strategy and then utilize them to solve the sample unbalancing and label unmatching between the source and target domain of WSOL. Experiments show that our pipeline outperforms SOTA methods on three WSOL benchmarks and can improve the performance of downstream weakly supervised semantic segmentation tasks.

Learning Domain-Invariant Model for WiFi-Based Indoor Localization

Learning Domain-Invariant Model for WiFi-Based Indoor Localization

Explainable artificial intelligence for traffic signal detection using LIME algorithm

As technology progresses, so does everything around us, such as televisions, mobile phones, and robots, which grow wiser. Of these technologies, artificial intelligence (AI) is used to aid the computer in making decisions comparable to humans, and this intelligence is supplied to the machine as a model. As AI deals with the concept of Black-Box, the model’s decisions were poorly comprehended by the end users. Explainable AI (XAI) is where humans can understand the judgments and decisions made by the AI. Earlier, the predictions made by the AI were not as easy as we know the data now, and there was some confusion regarding the predictions made by the AI. The intention for the use of XAI is to improve the user interface of products and services by helping them trust the decisions made by AI. The machine learning (ML) model White-box shows us the result that can be understood by the people in that domain, wherein the end users cannot understand the decisions. To further enhance traffic signal detection using XAI, the concept called local interpretable model- agnostic explanation (LIME) algorithm has been taken into consideration and the performance is improved in this paper.

Modelling of the trade-off between the deep soil moisture and vegetation restoration in the hilly area of the Loess Plateau, China

Excessive depletion of soil moisture by artificial forests in the vegetation restoration areas of the Loess Plateau has attracted widespread attention. To assess potential risks of soil moisture deficit, we needed on-site vegetation and soil sampling data, as well as UAV images from the Chaigou Watershed for three-dimensional analysis, combining both sampling and raster data. Three-dimensional surfaces for assessment of trade-off were established innovatively by the local regression and interpolation methods. The results indicated that soil moisture benefits at 20–40 cm depth are lower than at 0–20 cm due to infiltration and surface disturbance. In some areas of the Chaigou Watershed, grass and shrub-grass vegetation are facing risks of soil moisture deficit based on trade-off values (RSMD) and multiparameter evaluations. Analysis of deep soil water content variability revealed the moisture decreases significantly with the deepening of the soil layer in some plots. R (Richness), H (Shannon’s Diversity), Margalef, COHESION, and CONTAG were applicable in interpolation and fitted with the local regression model (R2 > 0.6) corresponding to the trade-off, but SPLIT was proven to be inapplicable in this study area. The zero trade-off inflection points were 6 %–8 %, while the trend inflection points were 7.5 %–9 % for soil moisture and vegetation indices of vegetation and landscape in typical sampling sites of the Chaigou Watershed. Three-dimensional fitting model is more comprehensive and effective in assessing deep soil moisture conditions and grass plots on shady slopes generally had the best trade-off status in this region.

Predicting coastal wave conditions: A simple machine learning approach

Accurate and reliable nearshore wave predictions are highly valuable for a range of marine activities, including coastal engineering and maritime transport. However, in nearshore locations, predicting wave properties is challenging due to complex shallow water processes, requiring local wave models. This article develops an alternative data-driven framework to predict wave parameters (e.g. significant wave height) through the extension of wave buoy datasets using a trained Gaussian process regression (GPR — a supervised machine learning method). We present an easy-to-implement workflow, where the extensive range of input parameters (from ECMWF’s (1) ERA5 reanalysis and (2) IFS forecast global wave model, ≈50km resolution) drives the development of GPR models. At five contrasting locations around the United Kingdom’s coastline, the GPR models produce wave predictions (forecast and hindcast) with low bias scores and strong correlations with observations. When compared to the global (ERA5 reanalysis) and a benchmark shelf-scale (Atlantic-European North West Shelf reanalysis; AENWS, 1.5−3.0km resolution) model, the GPR hindcasts reduced significant wave height (Hs) root-mean-squared error (RMSE) from 0.46 m (ERA5) and 0.21 m (AENWS) to 0.16 m (GPR). For the average zero-crossing wave period (Tz) RMSE reduced from 1.46 s (ERA5) and 1.15 s (AENWS) to 0.58 s (GPR). Because our approach uses publicly available global data, it can be implemented at any historic or active buoy location. We provide proof of concept for an online forecast and hindcast tool which has the potential to improve accessibility to coastal wave predictions for many marine stakeholders.

Local Metric NER: A new paradigm for named entity recognition from a multi-label perspective

As the field of Nested Named Entity Recognition (NNER) advances, it is marked by a growing complexity due to the increasing number of multi-label entity instances. How to more effectively identify multi-label entities and explore the correlation between labels is the focus of our work. Unlike previous models that are modeled in single-label multi-classification problems, we propose a novel multi-label local metric NER model to rethink Nested Entity Recognition from a multi-label perspective. Simultaneously, to address the significant sample imbalance problem commonly encountered in multi-label scenarios, we introduce a parts-of-speech-based strategy that significantly improves the model’s performance on imbalanced datasets. Experiments on nested, multi-label, and flat datasets verify the generalization and superiority of our model, with results surpassing the existing state-of-the-art (SOTA) on several multi-label and flat benchmarks. After a series of experimental analyses, we highlight the persistent challenges in the multi-label NER. We are hopeful that the insights derived from our work will not only provide new perspectives on the nested NER landscape but also contribute to the ongoing momentum necessary for advancing research in the field of multi-label NER.

Learning Critically: Selective Self-Distillation in Federated Learning on Non-IID Data

Federated learning (FL) enables multiple clients to collaboratively train a global model while keeping local data decentralized. Data heterogeneity (non-IID) across clients has imposed significant challenges to FL, which makes local models re-optimize towards their own local optima and forget the global knowledge, resulting in performance degradation and convergence slowdown. Many existing works have attempted to address the non-IID issue by adding an extra global-model-based regularizing item to the local training but without an adaption scheme, which is not efficient enough to achieve high performance with deep learning models. In this paper, we propose a Selective Self-Distillation method for Federated learning (FedSSD), which imposes adaptive constraints on the local updates by self-distilling the global model’s knowledge and selectively weighting it by evaluating the credibility at both the class and sample level. The convergence guarantee of FedSSD is theoretically analyzed and extensive experiments are conducted on three public benchmark datasets, which demonstrates that FedSSD achieves better generalization and robustness in fewer communication rounds, compared with other state-of-the-art FL methods.

Federated inception-multi-head attention models for cyber-attacks detection

<span lang="EN-US">With the proliferation of internet of things (IoT) devices, ensuring the security of these interconnected systems has become a critical concern. Cyberattacks targeting IoT devices pose significant threats to individuals and organizations due to the generation of vast amounts of data across many connected devices, which traditional centralized methods cannot solve. Federated learning (FL) could be a promising solution to mitigate privacy concerns associated with centralized approaches and address cybersecurity concerns. This paper uses FL and deep learning (DL) approaches to cybersecurity in IoT applications. The goal of cyber security is achieved by forming a federation of acquired and shared models at the head of the various participants. We use inception time and multi-head attention (CNN) algorithm based on FL to detect cyber-attacks and avoid data privacy leaks under two distribution modes, namely IID and Non-IID. In contrast, the FedAvg and FedMA algorithms aggregate local model updates. A global model is produced after several communication rounds between the IoT devices and the model parameter server. Cyber threats are simulated using edge-IIoT datasets. Experiment results show that the federated inception model's best global accuracy was 93, 91%, and 93, 49% using multi-head attention.</span>

Correlation-Embedded Transformer Tracking: A Single-Branch Framework.

Developing robust and discriminative appearance models has been a long-standing research challenge in visual object tracking. In the prevalent Siamese-based paradigm, the features extracted by the Siamese-like networks are often insufficient to model the tracked targets and distractor objects, thereby hindering them from being robust and discriminative simultaneously. While most Siamese trackers focus on designing robust correlation operations, we propose a novel single-branch tracking framework inspired by the transformer. Unlike the Siamese-like feature extraction, our tracker deeply embeds cross-image feature correlation in multiple layers of the feature network. By extensively matching the features of the two images through multiple layers, it can suppress non-target features, resulting in target-aware feature extraction. The output features can be directly used to predict target locations without additional correlation steps. Thus, we reformulate the two-branch Siamese tracking as a conceptually simple, fully transformer-based Single-Branch Tracking pipeline, dubbed SBT. After conducting an in-depth analysis of the SBT baseline, we summarize many effective design principles and propose an improved tracker dubbed SuperSBT. SuperSBT adopts a hierarchical architecture with a local modeling layer to enhance shallow-level features. A unified relation modeling is proposed to remove complex handcrafted layer pattern designs. SuperSBT is further improved by masked image modeling pre-training, integrating temporal modeling, and equipping with dedicated prediction heads. Thus, SuperSBT outperforms the SBT baseline by 4.7%,3.0%, and 4.5% AUC scores in LaSOT, TrackingNet, and GOT-10K. Notably, SuperSBT greatly raises the speed of SBT from 37 FPS to 81 FPS. Extensive experiments show that our method achieves superior results on eight VOT benchmarks.

Model-based dimensional NDE from few X-ray radiographs: Application to the evaluation of wall thickness in metallic turbine blades

The extraction of 3D dimensional measurements based on a limited number of 2D X-ray radiographs of a part would offer a significant speed-up of quality control procedures in industry. However, there are challenges with respect to both measurements and uncertainties. This work addresses these challenges by creating an estimated numerical model of the imaged part on which dimensional measurements can be made. The numerical model is chosen as a parametric deformable model that encodes the expected shape variability of the parts resulting from the manufacturing process. The parameters and uncertainties of the numerical model of the imaged part are estimated by the registration of the computed projections of the model and the observed radiographs without the need of any segmentation. The registration requires the model, the initial parameters, and the observed radiographs. The proposed approach is applied to the inspection of turbine blades manufactured by investment casting, and in particular to the measurement of their wall thickness, which is a critical control. The deformable model consists in partitioning the inner ceramic core into multiple subparts, which may undergo a rigid body motion with respect to the master die. Wall thickness measurements are determined from the estimation of these rigid body motions. To assess the reliability of the proposed procedure, a repeatability study is performed. In addition, wall thickness measurements were compared to corresponding measurements from the surface of the metal boundary obtained by X-ray computed tomography. This surface was determined from a reconstructed tomogram using commercial software. Both analyses show that such measurements are reliable and efficient. Furthermore, residual differences between captured and computed projections reveal localized shape deviations from the CAD model, meaning that despite localized model errors, the approach is operable.

Practical Attribute Reconstruction Attack Against Federated Learning

Existing federated learning (FL) designs have been shown to exhibit vulnerabilities which can be exploited by adversaries to compromise data privacy. However, most current works conduct attacks by leveraging gradients calculated on a small batch of data. This setting is not realistic as gradients are normally shared after at least 1 epoch of local training on each participant's local data in FL for communication efficiency. In this work, we conduct a unique systematic evaluation of attribute reconstruction attack (ARA) launched by the malicious server in the FL system, and empirically demonstrate that the shared local model gradients after 1 epoch of local training can still reveal sensitive attributes of local training data. To demonstrate this leakage, we develop a more effective and efficient gradient matching based method called cos-matching to reconstruct the sensitive attributes of any victim participant's training data. Based on the reconstructed training data attributes, we further show that an attacker can even reconstruct the sensitive attributes of any records that are not included in any participant's training data, thus opening a new attack surface in FL. Extensive experiments show that the proposed method achieves better attribute attack performance than existing state-of-the-art methods.

HDF-Net: Capturing Homogeny Difference Features to Localize the Tampered Image.

Modern image editing software enables anyone to alter the content of an image to deceive the public, which can pose a security hazard to personal privacy and public safety. The detection and localization of image tampering is becoming an urgent issue to be addressed. We have revealed that the tampered region exhibits homogenous differences (the changes in metadata organization form and organization structure of the image) from the real region after manipulations such as splicing, copy-move, and removal. Therefore, we propose a novel end-to-end network named HDF-Net to extract these homogeny difference features for precise localization of tampering artifacts. The HDF-Net is composed of RGB and SRM dual-stream networks, including three complementary modules, namely the suspicious tampering-artifact prominent (STP) module, the fine tampering-artifact salient (FTS) module, and the tampering-artifact edge refined (TER) module. We utilize the fully attentional block (FLA) to enhance the characterization ability of homogeny difference features extracted by each module and preserve the specifics of tampering artifacts. These modules are gradually merged according to the strategy of "coarse-fine-finer", which significantly improves the localization accuracy and edge refinement. Extensive experiments demonstrate that HDF-Net performs better than state-of-the-art tampering localization models on five benchmarks, achieving satisfactory generalization and robustness.

SBPA: Sybil-Based Backdoor Poisoning Attacks for Distributed Big Data in AIoT-Based Federated Learning System

Federated learning (FL) enables a great deal of distributed independent participants to collaborate in training without sharing data. Malicious adversary can poison the local model by backdoor poisoning attacks and utilize the characteristic that server cannot trace original data to make the poisoned model directly aggregated. Especially in the AIoT-FL network that generates large amounts of data in real-time, such an attack is more powerful. In this paper, we design a sybil-based backdoor poisoning attacks (SBPA) against the above vulnerability. Malicious participants inject backdoor triggers into distributed Big Data to covertly complete data poisoning. During subsequent iterative aggregation, the joint model activates the backdoor during testing to achieve misclassification for the backdoor images. Besides, malicious participants create some sybil nodes to join the aggregation by taking advantage of the vulnerability of system devices to be easily disconnected. They are committed to making the poisoned local models aggregated with higher probability. Their goal works on making the final global model misclassify backdoor images while keeping high classification accuracy on the other non-backdoor samples. We conduct extensive experiments on multiple datasets and exhibit more robust performance than the state-of-the-art in various metrics under both data distribution including i.i.d. and non-i.i.d. scenarios.

Application of an Improved State Feedback Control in the Selective Catalytic Reduction Denitrification Systems of Coal Power Units Under Variable Load Conditions

Selective catalytic reduction (SCR) flue gas denitrification systems are inherently complex, typically embodying characteristics of non-linearity, significant time delays, and susceptibility to multiple disturbances. In the context of coal power units engaging in deep load cycling and rapid frequency adjustment, conventional proportional-integral-derivative (PID) control struggles to meet the demands of effective control. This study introduces a control strategy that incorporates a “state observer + Linear Quadratic Regulator (LQR) state feedback + Improved Quantum Genetic Algorithm (IQGA) optimized PID”. Initially, local linear mathematical models of an SCR denitrification system at 340 MW, 450 MW, and 540 MW loads were used to design state observer and LQR state feedback control parameters for each operational condition. At a single load point, the IQGA was employed to optimize the outer loop PID parameters, followed by simulation experiments of load increases and decreases between 340 MW and 540 MW. The results demonstrated that, compared to two other strategies, the proposed approach reduced the overshoot by a minimum of 1.5% and shortened the adjustment time by 31.7% under conditions of step disturbances and internal perturbations. Throughout variable operational conditions, the strategy consistently exhibited minimal output fluctuations, rapid adjustment capabilities, strong disturbance rejection, and robust stability. This algorithm proves to be an effective method for controlling NOx concentrations, offering insights for precise ammonia injection control in future applications.

Evaluating the Efficacy of LLMs to Emulate Realistic Human Personalities

To enhance immersion and engagement in video games, the design of Affective Non-Player Characters (NPCs) is a key focus for researchers and practitioners. Affective Computing frameworks improve Non-player characters (NPC) by providing personalities, emotions, and social relations. Large Language Models (LLMs) bring the promise to dynamically enhance character design when coupled with these frameworks, but further research is needed to validate the models truly represent human qualities. In this research, a comprehensive analysis investigates the capabilities of LLMs to generate content that aligns with human personality, using the Big Five and human responses from the International Personality Item Pool (IPIP) questionnaire. Our goal is to benchmark the performance of various LLMs, including frontier models and local models, against an extensive dataset comprising over 50,000 human surveys of self-reported personality tests to determine whether LLMs can replicate human-like decision-making with personality-driven prompts. A range of personality profiles were used to cluster the test results from the human survey dataset. Our methodology involved prompting LLMs with self-evaluated test items for each personality profile, comparing their outputs to human baseline responses, and evaluating the accuracy and consistency. Our findings show that some local models had 0% alignment of any personality profiles when compared to the human dataset, while the frontier models, in some cases, had 100% alignment. The results indicate that NPCs can successfully emulate human-like personality traits using LLMs, as demonstrated by benchmarking the LLM's output against human data. This foundational work serves as a methodology for game developers and researchers to test and evaluate LLMs, ensuring they accurately represent the desired human personalities and can be expanded for further validation.

Measurements of N2 + ions in an inductively coupled plasma using saturated cavity ringdown spectroscopy

Abstract This paper presents a unique study of the bulk plasma characteristics in a low pressure inductively coupled nitrogen plasma. Saturated cavity ringdown spectroscopy (sat-CRDS) has been used to determine the absolute number densities and translational temperatures of N2 +(X, ν = 0). The effect of saturation is readily accounted for by using an effective saturation parameter, Seff , and determined by a simple method employing measurements at two different gain settings of the detection system. The appropriateness of this method is confirmed by comparison with fitting individual ringdown data using a time-dependent saturation parameter, S(t), within the local approximation model for sat-CRDS; the two methods are in excellent agreement in returning absolute number densities and translational temperatures. N2 +(X, ν = 0) number densities are determined across a matrix of pressure (10 − 100 mTorr) and rf power (200 − 400 W) conditions with maximum number densities of ca. 1.3 × 1010 cm−3 while translational temperatures range from 600 − 1500 K.

High-Order Local Clustering on Hypergraphs

Graphs are a commonly used model in data mining to represent complex relationships, with nodes representing entities and edges representing relationships. However, graphs have limitations in modeling high-order relationships. In contrast, hypergraphs offer a more versatile representation, allowing edges to join any number of nodes. This capability empowers hypergraphs to model multiple relationships and capture high-order information present in real-world applications. We focus on the problem of local clustering in hypergraphs, which computes a cluster near a given seed node. Although extensively explored in the context of graphs, this problem has received less attention for hypergraphs. Current methods often directly extend graph-based local clustering to hypergraphs, overlooking their inherent high-order features and resulting in low-quality local clusters. To address this, we propose an effective hypergraph local clustering model. This model introduces a novel conductance measurement that leverages the high-order properties of hypergraphs to assess cluster quality. Based on this new definition of hypergraph conductance, we propose a greedy algorithm to find local clusters in real time. Experimental evaluations and case studies on real-world datasets demonstrate the effectiveness of the proposed methods.