Anomaly detection plays a vital role in processing multi-source data through public cloud servers, yet existing privacy-preserving schemes fail to efficiently detect anomalies while protecting data source privacy. Although isolation forest offer advantages for unsupervised high-dimensional data analysis, implementing its tree-based privacy-preserving mechanisms remains challenging. In this paper, we propose IFAD, a novel isolation forest-based scheme for detecting anomalies in private data. IFAD guarantees end-to-end privacy protection by safeguarding original data, tree structures, and intermediate information throughout detection workflows. Our design achieves efficiency through three key contributions: 1) Cryptographic building blocks combining function secret sharing (FSS) and secret sharing (SS) to enable secure computations; 2) A split index protocol and layer update protocol to facilitate efficient, layer-by-layer isolation forest construction; 3) A detection phase optimization converting the anomaly score calculations into lookup table operations. Experimental evaluations demonstrate that IFAD achieves superior performance, outperforming prior schemes by 2.4×-3.1× in runtime under LAN and WAN environments, and by 1.8×-7.8× in online communication overhead, while maintaining comparable detection accuracy. Our solution establishes an effective balance between privacy preservation and operational efficiency for cloud-based anomaly detection.

Principal component analysis (PCA) is an essential algorithm for dimensionality reduction in various data analysis tasks. Recently, PCA has gained widespread use in cloud outsourcing services due to its effectiveness and versatility. However, privacy concerns in outsourced PCA have led to the development of privacy-preserving schemes. Despite this, existing solutions face significant performance bottlenecks due to the iterative matrix computations involved in PCA, resulting in high overhead that limits their practicality. In this paper, we propose an efficient privacy-preserving outsourced PCA scheme. Specifically, we propose a secure Jacobi-EVD protocol, which improves efficiency by reducing nonlinear operations and iterations. Furthermore, by optimizing the matrix update operator in Jacobi-EVD using a hybrid protocol, we significantly reduce the communication overhead and communication rounds in iterative matrix computations. Security analysis demonstrates that our scheme preserves the privacy of data and PCA results. Performance evaluation shows our scheme significantly reduces 29.3× communication overhead compared to existing schemes.

Federated learning addresses the issue of data silo in machine learning. However, in practical applications, it still encounters challenges such as privacy leakage and communication bottleneck. Previous studies have proposed two main technologies to these challenges: secure aggregation to preserve privacy and Top-k gradient sparsification to reduce communication overhead, respectively. However, for both privacy preservation and communication efficiency, combining these two technologies results in compatibility issues and additional privacy leakage. In this paper, we propose a secure aggregation protocol with Top-k sparsification to achieve secure and efficient federated learning. We employ a differential privacy perturbation mechanism to protect Top-k features, thus preventing client's privacy leakage. Additionally, we design a sparse communication graph to ensure compatibility between secure aggregation and Top-k sparsification perturbed by differential privacy. We prove that our protocol protects Top-k features and conduct extensive experiments to evaluate its performance, which shows a significant reduction in the communication overhead compared to traditional secure aggregation protocols.

Automated mechanistic interpretation research has attracted great interest due to its potential to scale explanations of neural network internals to large models. Existing automated circuit discovery work relies on activation patching or its approximations to identify subgraphs in models for specific tasks (circuits). They often suffer from slow runtime, approximation errors, and specific requirements of metrics, such as non-zero gradients. In this work, we introduce contextual decomposition for transformers (CD-T) to build interpretable circuits in large language models. CD-T can produce circuits at any level of abstraction and is the first to efficiently produce circuits as fine-grained as attention heads at specific sequence positions. CD-T is compatible to all transformer types, and requires no training or manually-crafted examples. CD-T consists of a set of mathematical equations to isolate contribution of model features. Through recursively computing contribution of all nodes in a computational graph of a model using CD-T followed by pruning, we are able to reduce circuit discovery runtime from hours to seconds compared to state-of-the-art baselines. On three standard circuit evaluation datasets (indirect object identification, greater-than comparisons, and docstring completion), we demonstrate that CD-T outperforms ACDC and EAP by better recovering the manual circuits with an average of 97% ROC AUC under low runtimes. In addition, we provide evidence that faithfulness of CD-T circuits is not due to random chance by showing our circuits are 80% more faithful than random circuits of up to 60% of the original model size. Finally, we show CD-T circuits are able to perfectly replicate original models' behavior(faithfulness = 1) using fewer nodes than the baselines for all tasks. Our results underscore the great promise of CD-T for efficient automated mechanistic interpretability, paving the way for new insights into the workings of large language models.

Room layout estimation from multiple-perspective images is poorly investigated due to the complexities that emerge from multi-view geometry, which requires muti-step solutions such as camera intrinsic and extrinsic estimation, image matching, and triangulation. However, in 3D reconstruction, the advancement of recent 3D foundation models such as DUSt3R has shifted the paradigm from the traditional multi-step structure-from-motion process to an end-to-end single-step approach. To this end, we introduce Plane-DUSt3R, a novel method for multi-view room layout estimation leveraging the 3D foundation model DUSt3R. Plane-DUSt3R incorporates the DUSt3R framework and fine-tunes on a room layout dataset (Structure3D) with a modified objective to estimate structural planes. By generating uniform and parsimonious results, Plane-DUSt3R enables room layout estimation with only a single post-processing step and 2D detection results. Unlike previous methods that rely on single-perspective or panorama image, Plane-DUSt3R extends the setting to handle multiple-perspective images. Moreover, it offers a streamlined, end-to-end solution that simplifies the process and reduces error accumulation. Experimental results demonstrate that Plane-DUSt3R not only outperforms state-of-the-art methods on the synthetic dataset but also proves robust and effective on in the wild data with different image styles such as cartoon. Our code is available at: https://github.com/justacar/Plane-DUSt3R

Mobile users often prefer to sense only a subset of tasks based on their preferences or physical conditions, which distinguishes sparse crowdsensing from traditional crowdsensing. Sparse crowdsensing not only introduces a potential risk of privacy leakage regarding users’ preferences or conditions—due to the revelation of specific sensed objects—but also results in reduced accuracy of truth estimation. To address these challenges, we propose a Privacy-Preserving Truth Discovery (PPTD) scheme, named S-PPTD, that enables accurate and efficient PPTD for sparse crowdsensing. Our approach leverages edge nodes to geographically group users and introduces an effective padding strategy based on Bloom filters and mixed secret sharing. This strategy allows users to obfuscate the objects they sense, preventing adversaries from determining the specific objects being sensed. To improve accuracy, we design new protocols for precise and efficient approximation of nonlinear functions, enabling the use of commonly applied kernel functions to capture spatial and temporal correlations between objects, and incorporate these into the truth estimation process. Through extensive experiments and security analysis, we demonstrate that S-PPTD is secure, accurate, and efficient in the context of sparse mobile crowdsensing.

For shipboard coherent S-band radar, the shadow modulation phenomenon happens at large sea surface undulations or small grazing angles. As the ship speed increases or the sea state rises, the shadow modulation leads to an increase in the peak value of the wave spectrum, which results in inaccurate wave parameter inversion. In order to achieve accurate wave detection in high sea state or high ship speed scenarios, a shadow modulation correction method based on velocity-fitting relationship and Hermite interpolation is presented. First, the variable range of radial velocity is computed using the simulated sea surface’s velocity fitting relation. The anomalous Doppler velocity is then rectified by applying the modified Akima piecewise cubic Hermite interpolation (MAPCHI) algorithm. The MAPCHI algorithm is appropriate for complex nonlinear curves and is not prone to abrupt fluctuations or flattening issues. The implementation of this interpolation algorithm can alleviate the problem of anomalous fluctuations in Doppler velocities induced by shadow modulation. Finally, the nondirectional wave spectrum and wave parameters are calculated. Numerical simulation experiments show that the proposed method can rectify the influence of shadow modulation at various speeds and sea state backdrops. Meanwhile, comparing the radar measurements in the South China Sea with the buoy results, the mean absolute error (MAE) of significant wave height and mean wave period are 0.21 m and 0.34 s, respectively. The results demonstrate that the proposed method can successfully remove shadow modulation interference with high robustness.

Set-valued data are commonly used to represent subsets of a universal set and are frequently utilized in online services, such as online shopping preferences, website browsing records, and recently visited places. By collecting set-valued data from users, service providers can perform statistical analysis to obtain a joint distribution of service usage data and subsequently learn the association between different kinds of set-valued data to improve the quality of service. However, collecting set-valued data raises privacy concerns about the potential misuse of records to infer individuals’ identities and preferences. Although some privacy-preserving aggregation mechanisms for set-valued data have been proposed, they have not yet achieved joint distribution analysis with high accuracy. In this paper, we propose a joint distribution analysis method for set-valued data with local differential privacy (LDP). We design a scalable perturbation mechanism under$\epsilon $-LDP by limiting the range of users’ responses in the collection process and cyclically shifting the set-valued data in an encoded uniform format, ensuring that the size of the universal set does not influence the accuracy of the results. Based on the perturbation method, we develop an analysis method to efficiently obtain association information between two sets. By performing specific bitwise operations on the perturbed data matrices, the computational overhead is linear with respect to the cardinality of the item set. In addition to theoretically analyzing the error bound and proving the security of our work, extensive experimental results on synthetic and real-world datasets demonstrate that our scheme achieves better utility than existing state-of-the-art approaches.

The partial ordered data is typically used to describe the order of some elements within a set, and it widely exists in various fields, such as clinical investigations, preference ranking and voting. However, the collection of partial ordered data poses critical privacy concerns about abusing records to infer individuals’ identities and preferences. To solve this problem, this paper proposes a distribution analysis method for partial ordered data with local differential privacy (LDP). The private information of partial ordered data includes whether an element is associated with a partial order relation and either a relation is preceding or succeeding. To preserve privacy, we perturb partial ordered data by randomly responding raw data or the data with mapped elements. This makes it impossible to distinguish whether any element has a partial order relationship with other elements and what kind of partial order relationship exists. To maintain the logicality of partial ordered data, we utilize the transitivity of partial orders to distinguish between direct and indirect orders in the perturbation. The inherent properties of partial orders are still satisfied after perturbation, which reduces the possibility of servers inferring the raw data through logical errors. Moreover, we theoretically analyze the error bound and prove the security of our work. Extensive experimental results on synthetic and real-world datasets demonstrate that our scheme achieves better utility than existing state-of-the-art approaches.

Truth discovery is an effective tool to infer true information from multi-source data and has been widely applied in mobile crowdsensing systems. In some specific scenarios, the sensory data are collected in a streaming fashion with time-varying information, and the server should update the truth in time. Under such circumstances, local differential privacy-based mechanism can satisfy the requirement of real-time processing properly while keeping the privacy of sensory data. However, directly applying local differential privacy to handle streaming data will disclose the long-term potential privacy and decrease the accuracy. To address these problems, we propose an incentive-based privacy-preserving truth discovery framework over streaming data. Firstly, we adopt the sequential composition theorem of w-event privacy to protect workers' long-term privacy. Second, we design an incentive mechanism to improve the submitted data utility and thus avoid the decrease in accuracy. In this way, our scheme ensures that workers submit more accurate data while their global privacy is still guaranteed. Finally, we prove our scheme satisfies w-event (∊, δ) differential privacy and theoretically analyze the result utility. Extensive experiments also demonstrate the effectiveness of our incentive mechanism.