As on-device large language model (LLM) systems become increasingly prevalent, federated fine-tuning enables advanced language understanding and generation directly on edge devices; however, it also involves processing sensitive, user-specific data, raising significant privacy concerns within the federated learning framework. To address these challenges, we propose DP-FedLoRA, a privacy-enhanced federated fine-tuning framework that integrates LoRA-based adaptation with differential privacy in a communication-efficient setting. Each client locally clips and perturbs its LoRA matrices using Gaussian noise to satisfy (∊, δ)-differential privacy. We further provide a theoretical analysis demonstrating the unbiased nature of the updates and deriving bounds on the variance introduced by noise, offering practical guidance for privacy-budget calibration. Experimental results across mainstream benchmarks show that DP-FedLoRA delivers competitive performance while offering strong privacy guarantees, paving the way for scalable and privacy-preserving LLM deployment in on-device environments.

Current recommendation systems recommend goods by considering users' historical behaviors, social relations, ratings, and other multi-modals. Although outdated user information presents the trends of a user's interests, no recommendation system can know the users' real-time thoughts indeed. With the development of brain-computer interfaces, it is time to explore next-generation recommenders that show users' real-time thoughts without delay. Electroencephalography (EEG) is a promising method of collecting brain signals because of its convenience and mobility. Currently, there is only few research on EEG-based recommendations due to the complexity of learning human brain activity. To explore the utility of EEG-based recommendation, we propose a novel neural network model, QUARK, combining Quantum Cognition Theory and Graph Convolutional Networks for accurate item recommendations. Compared with the state-of-the-art recommendation models, the superiority of QUARK is confirmed via extensive experiments.

When multimodal AI systems increasingly utilize diverse data sources to achieve advanced understanding and interaction, they inevitably collect vast amounts of sensitive information, thus highlighting the urgent need for robust privacy safeguards, especially as these technologies expand into fields like healthcare, finance, and education. Existing research on data privacy in AI, encompassing adversarial training-based models, differential privacy-based models, and differentially private transform-based models, often neglects the inter-correlation inherent in multi-sensor data. To address this gap, we propose the differentiAl Private OnLine muLti-sensor data predictiOn model (APOLLO), which simultaneously considers intra-correlation and inter-correlation to enhance privacy protection while maintaining predictive performance. Under the proposed APOLLO frame-work, we design two implementations: APOLLO I, which ensures$\epsilon$-differential privacy by adding Laplace noise to each correlated data segment, and APOLLO II, which applies additional noise to make the concatenated multi-sensor data realize$\epsilon{-}$differential privacy. Furthermore, we conduct the theoretical analysis to reveal the relationship between performance influence and the privacy budget, providing guidelines for noise addition with the aim of achieving certified performance. Comprehensive experiments validate the effectiveness of the APOLLO model, establishing a new standard for privacy-preserving multi-sensor data prediction.

Machine unlearning is an emerging need that aims to remove the influence of deleted data from a learned model in a timely manner. Thus, unlearning is important for privacy and security in data management. Nevertheless, existing machine unlearning methods fail to perform exactly and efficiently in a federated setting. In this paper, we study the unlearning problem in federated learning, which provides a data deletion mechanism in the federated setting. First of all, a quantized federated learning (Q-FL) algorithm is developed to facilitate exact unlearning. Based on the quantized federated learning system, an exact and efficient federated unlearning (Exact-Fun) algorithm is designed to realize the goal of data deletion. Through theoretic analysis and experimental evaluation, our proposed methods not only have the desired unlearning effectiveness but also achieve high unlearning efficiency compared with the existing works.

3D grey image segmentation has become a promising approach to facilitate practical applications with the help of advanced deep learning models. Although a number of previous works have investigated the vulnerability of deep learning models to backdoor attack, there is no work to study the severe risk of backdoor attack on 3D grey image segmentation. To this end, we propose two backdoor attack methods on 3D grey image segmentation, including Full-control Backdoor Attack (FCBA) and Partial-control Backdoor Attack (PCBA), on 3D grey image segmentation by leveraging a frequency trigger injection function and a rotation-based label corruption function. Our proposed trigger injection function is applied to insert a 3D trigger pattern into the benign 3D grey images in the frequency domain while ensuring the invisibility of the trigger pattern. And the proposed rotation-based label corruption function is employed to yield the crafted labels with the aim of decreasing the performance of segmentation. Finally, through comprehensive experiments on a real-world dataset, we demonstrate the effectiveness of our proposed backdoor models, the frequency trigger injection function, and the rotation-based label corruption function.

Analyzing dynamic information networks, which contain evolving objects and links, to meet users various needs has attracted much attention in recent years. For sales companies, recruiting staff who are socialites would help to increase sales volume since such staff often sell more. For universities, employing active collaborators who are productive and well connected to many different scholars over time could bring many benefits to their development. However, no previous work has focused on the discovery of socialites and active collaborators who are worthy of investment in reality. In this paper, we advocate a new concept of attractive individuals to model such special objects. We also introduce the concept of attractive groups to represent groups of well-connected attractive individuals. We analyze the complexity of the attractive individual and group search problems. A time and space efficient algorithm is presented to detect attractive individuals. Furthermore, three algorithms are respectively proposed to find top-k representative attractive groups. Experiments on 6 real-world datasets show high performance of our methods and the significance of attractive individuals and groups in reality.

Searching for$k$-cliques in graphs has been an important problem in graph analysis due to its large number of applications. Previously, finding$k$-cliques in weighted graphs aimed at finding cliques with the largest sum of weight (with no distinction between the edge or the vertex weights), usually called the sum model. However, the algorithms under the sum model may result in solutions consisting of low-weight vertices or edges (outliers). To address this issue, we propose a new model named maximal$(S, C, K)$-clique in weighted graphs and study the problem of maximal ($S, C, K$)-clique search (MCS). We first propose an enumeration-based algorithm MCSE, which checks every$k$-clique to identify the maximal ($S, C, K$)-clique. To improve the efficiency, we further propose two improved algorithms MCSP and MCSC. Instead of checking every possible$k$-clique, MCSP focuses on ($S, C$) values that cannot be dominated and obtains the maximal$(S, C, K)$-cliques directly based on these values. MCSC is devised by further optimizing MCSP based on some key observations on maximal cliques and cliques’ nesting property. We also propose two index structures, BCS-Index and ICS-Index, to achieve optimal query. The former stores all maximal$(S, C, K)$-cliques, while the latter uses the clique's nesting property to reduce the space cost of index construction. Extensive experiments conducted on six real graphs demonstrate the efficiency and effectiveness of our proposed algorithms.

Micro-video recommendation is attracting global attention and becoming a popular daily service for people of all ages. Recently, Graph Neural Networks-based micro-video recommendation has displayed performance improvement for many kinds of recommendation tasks. However, the existing works fail to fully consider the characteristics of micro-videos, such as the high timeliness of news nature micro-video recommendation and sequential interactions of frequently changed interests. In this paper, a novel Multi-aggregator Time-warping Heterogeneous Graph Neural Network (MTHGNN) is proposed for personalized news nature micro-video recommendation based on sequential sessions, where characteristics of micro-videos are comprehensively studied, users' preference is mined via multi-aggregator, the temporal and dynamic changes of users' preference are captured, and timeliness is considered. Through the comparison with the state-of-the-arts, the experimental results validate the superiority of our MTHGNN model.

Analyzing dynamic information networks, which contain evolving objects and links, has attracted much attention in recent years. For sales companies, recruiting staff who are socialites would help to increase sales volume since such staff often sell more. For universities, employing active collaborators who are productive and well connected to many different scholars over time could bring many benefits to their development. However, no previous work has focused on socialites and active collaborators who are worthy of investment in reality. In this paper, we advocate attractive individuals to model such special objects, and introduce attractive groups to represent groups of well-connected attractive individuals. Then we propose several algorithms to search attractive individuals and groups. Extensive experiments show high performance of our methods and the significance of attractive individuals and groups in reality.

Multi-label image recognition has been an indispensable fundamental component for many real computer vision applications. However, a severe threat of privacy leakage in multi-label image recognition has been overlooked by existing studies. To fill this gap, two privacy-preserving models, Privacy-Preserving Multi-label Graph Convolutional Networks (P2-ML-GCN) and Robust P2-ML-GCN (RP2-ML-GCN), are developed in this article, where differential privacy mechanism is implemented on the model’s outputs so as to defend black-box attack and avoid large aggregated noise simultaneously. In particular, a regularization term is exploited in the loss function of RP2-ML-GCN to increase the model prediction accuracy and robustness. After that, a proper differential privacy mechanism is designed with the intention of decreasing the bias of loss function in P2-ML-GCN and increasing prediction accuracy. Besides, we analyze that a bounded global sensitivity can mitigate excessive noise’s side effect and obtain a performance improvement for multi-label image recognition in our models. Theoretical proof shows that our two models can guarantee differential privacy for model’s outputs, weights and input features while preserving model robustness. Finally, comprehensive experiments are conducted to validate the advantages of our proposed models, including the implementation of differential privacy on model’s outputs, the incorporation of regularization term into loss function, and the adoption of bounded global sensitivity for multi-label image recognition.

In this paper, we study exact hypercore maintenance in large-scale dynamic hypergraphs. A hypergraph, whose hyperedges may contain a set of vertices rather than two vertices in pairwise graphs, can represent complex interactions in more sophisticated applications. However, the exponential number of hyperedges incurs unaffordable costs to recompute the hypercore number of vertices and hyperedges when updating a hypergraph. This motivates us to propose an efficient approach for exact hypercore maintenance with the intention of significantly reducing the hypercore updating time comparing with recomputation approaches. The proposed algorithms can pinpoint the vertices and hyperedges whose hypercore numbers have to be updated by only traversing a small sub-hypergraph. Extensive experiments on real-world and temporal hypergraphs demonstrate the superiority of our algorithms in terms of efficiency.

The increasing of group polarization on social media seriously impacts on the health of public discourse and information dissemination. At present, detecting polarized structures in signed networks is well-motivated for studying the group polarization on social media. However, most studies restricted the number of polarized structures to only two, while neglecting the real-world scenario where signed networks consist of multiple polarized structures, that is an unreasonable assumption. To conquer the limitations of the existing work, in this paper, we present a novel cohesive subgraph model based on structural clusterable theory, named maximal multipolarized clique (MMC), which can be partitioned into k polarized subcliques such that the edges in subcliques are positive and the edges between subcliques are negative. This paper formulates the problem of Maximal Multipolarized Cliques Search (MMCS) in signed networks which is proved to be NP-hard. To address this problem, we first devise powerful pruning rules to reduce the signed network significantly and further develop an efficient algorithm to search all maximal multipolarized cliques in the reduced signed network. The experimental results on real-world signed networks demonstrate the efficiency and effectiveness of our algorithm.

With the rise of the Internet of Things (IoT) and fifth-generation (5G) networks, which have led to a surge in data processing and increased data transfer time, traditional cloud computing could no longer meet the needs of workers, so edge computing has emerged. Edge computing could meet the demand for low time consumption by processing data at the edge of the network and then transmitting it to a third-party platform. However, since the credibility of the third-party platform is unknown which can easily leak the privacy of workers. For the transparent mechanism of blockchain, a two-stage privacy protection mechanism based on blockchain is proposed to solve this problem. In the first stage, this paper proposes a double disturbance localized differential privacy (DDLDP) algorithm to disturb the location information of workers. In the second stage, all the sensing data are uploaded to the blockchain through edge nodes, processed by the edge cloud, and fed back to the requester. Blockchain technology not only guarantees the integrity of sensing data, but also prevents the possibility of third-party platforms from leaking workers' privacy. Through extensive performance evaluation and comparative experiments on real data sets, the DDLDP algorithm could effectively protect the privacy of workers and has higher service quality and data availability.

The huge amount of graph data are published and shared for research and business purposes, which brings great benefit for our society. However, user privacy is badly undermined even though user identity can be anonymized. Graph de-anonymization to identify nodes from an anonymized graph is widely adopted to evaluate users' privacy risks. Most existing de-anonymization methods which are heavily reliant on side information (e.g., seeds, user profiles, community labels) are unrealistic due to the difficulty of collecting this side information. A few graph de-anonymization methods only using structural information, called seed-free methods, have been proposed recently, which mainly take advantage of the local and manual features of nodes while overlooking the global structural information of the graph for de-anonymization.

The skyline query is important in the database community. In recent years, the researches on incomplete data have been increasingly considered, especially for the skyline query. However, the existing skyline definition on incomplete data cannot provide users with valuable references. In this paper, we propose a novel skyline definition utilizing probabilistic model on incomplete data where each point has a probability to be in the skyline. In particular, it returns K points with the highest skyline probabilities. In addition, we propose incomplete models and estimate probability density functions of missing values on independent, correlated, and anti-correlated distributions, respectively. Meanwhile, it is a big challenge to compute probabilistic skyline on incomplete data. We propose three efficient algorithms SPISkyline, SPCSkyline, and SPASkyline for probabilistic skyline computation on incomplete data complying with independent, correlated, and anti-correlated distributions, respectively. They employ pruning strategy, optimization of the process of probability computation, and sorting technique to improve the efficiency of probabilistic skyline computation on incomplete data. Our experimental results demonstrate that our proposed concept of probabilistic skyline is an effective method to tackle skyline query on incomplete data and our algorithms are tens of times faster than the naive algorithm on both synthetic and real datasets.

This paper studies the deletion propagation problem in terms of minimizing view side-effect. It is a problem funda-mental to data lineage and quality management which could be a key step in analyzing view propagation and repairing data. The investigated problem is a variant of the standard deletion propagation problem, where given a source database D, a set of key preserving conjunctive queries Q, and the set of views V obtained by the queries in Q, we try to identify a set T of tuples from D whose elimination prevents all the tuples in a given set of deletions on views △V while preserving any other results. The complexity of this problem has been well studied for the case with only a single query. Dichotomies, even trichotomies, for different settings are developed. However, no results on multiple queries are given which is a more realistic case. We study the complexity and approximations of optimizing the side-effect on the views, i.e., find T to minimize the additional damage on V after removing all the tuples of △V. We focus on the class of key-preserving conjunctive queries which is a dichotomy for the single query case. It is surprising to find that except the single query case, this problem is NP-hard to approximate within any constant even for a non-trivial set of multiple project-free conjunctive queries in terms of view side-effect. The proposed algorithm shows that it can be approximated within a bound depending on the number of tuples of both V and △V. We identify a class of polynomial tractable inputs, and provide a dynamic programming algorithm to solve the problem. Besides data lineage, study on this problem could also provide important foundations for the computational issues in data repairing. Furthermore, we introduce some related applications of this problem, especially for query feedback based data cleaning.

The autonomous driving (auto-driving) technology has been promoted significantly by the rapid advances in computer vision and deep neural networks. Auto-driving vehicles, nowadays, are fully equipped with numerous sensors such as cameras, geo-sensors, and radar sensors, to capture real-time data inside the vehicles and outside surroundings. Meanwhile, the captured data contains lots of private information about vehicles, drivers and passengers and thus faces a high risk of privacy breaches. Especially, side-channel information can be mined from camera data to identify vehicles' locations and even trajectories, raising serious privacy issues. Unfortunately, the issue, how to resist location-inference attack for camera data in auto-driving, has never been addressed in literature. In this paper, we intend to fill this blank by developing a GAN-based image-toimage translation method named Auto-Driving GAN (ADGAN). Through performance comparisons between ADGAN and the state-of-the-art, the superiority of ADGAN can be validated - offering an effective tradeoff between recognition utility and privacy protection for camera data.

The view propagation problem is a class of view update problem in relational databases [7], involving deletion and insertion propagations. Given source database D, conjunctive query Q, view V generated by query Q(D) and a deletion (insertion) on view ΔV, deletion (insertion) propagation is to find a side effect free update ΔD on D such that the deletion (insertion) of ΔD from (into) D will delete (insert) the intentional ones ΔV without resulting in the deletion (insertion) of additional tuples from (into) the view. Generally, such a deletion (insertion) is side effect free. The related data management applications include query result explanation, data debugging, and anonymizing datasets, which rely on understanding how interventions in a database affect the output of a query. View propagation is a natural and typical way to define such interventions, which seems to be well-studied. However, in general, the candidate update on a source database is picked up aimlessly in advance, making the updated database to be very distant from the original one no matter whether it is the maximum one. In this paper, we formally define the bounded view propagation problem, where candidate update ΔD is bounded as a subset of potential C which is a fixed small tuple set of D. We study the complexity of this problem for conjunctive queries, and make contributions to the previous results of the problems of side-effect free deletion propagation. Specifically, our bounded view propagation problem decreases computational complexity regardless of conjunctive query structure. We show the fixed potential is actually a dichotomy for both deletion and insertion propagations, and figure out the results on combined complexity which is neglected previously. Based on our results, for view propagation, we map out a complete picture of the computational complexity hierarchy for conjunctive queries on both data and combined complexities. Moreover, this bounded version is an update forbidden case of view propagation, and our results can be applied to it.

The skyline query is important in database community. In recent years, the researches on incomplete data have been increasingly considered, especially for the skyline query. However, the existing skyline definition on incomplete data cannot provide users with valuable references. In this paper, we propose a novel skyline definition utilizing probabilistic model on incomplete data where each point has a probability to be in the skyline. In particular, it returnsK points with the highest skyline probabilities. Meanwhile, it is a big challenge to compute probabilistic skyline on incomplete data. We propose an efficient algorithm PISkyline, which utilizes two pruning strategies to reduce the number of points and adopts two optimizations to accelerate probability computation for each point. Nevertheless, PISkyline is susceptible to the order of input data and there is still a great deal of room for optimization. We develop a point-level sorting technique by adjusting the order of accessing points to further improve the efficiency of PISkyline. Our experimental results demonstrate that our algorithms are tens of times faster than the naive algorithm on both synthetic and real datasets.

The amount of sensory data manifests an explosive growth due to the increasing popularity of Wireless Sensor Networks (WSNs). The scale of sensory data in many applications has already exceeded several petabytes annually, which is beyond the computation and transmission capabilities of conventional WSNs. On the other hand, the information carried by big sensory data has high redundancy because of strong correlation among sensory data. In this paper, we introduce the novel concept of ϵ-Kernel Dataset, which is only a small data subset and can represent the vast information carried by big sensory data with the information loss rate being less than ϵ, where ϵ can be arbitrarily small. We prove that drawing the minimum ϵ-Kernel Dataset is polynomial time solvable and provide a centralized algorithm with O(n3) time complexity. Furthermore, a distributed algorithm with constant complexity O(1) is designed. It is shown that the result returned by the distributed algorithm can satisfy the ϵ requirement with a near optimal size. Furthermore, two distributed algorithms of maintaining the correlation coefficients among sensor nodes are developed. Finally, the extensive real experiment results and simulation results are presented. The results indicate that all the proposed algorithms have high performance in terms of accuracy and energy efficiency.