Advancements in large language models (LLMs) are driving progress towards general artificial intelligence. Ensuring differential privacy during fine-tuning and inference on domain-specific data is essential, but balancing privacy and utility, especially for sequential and label data, remains challenging. Existing methods protect sequential data well but struggle with label data due to traditional definitions of label differential privacy, which ignore label-feature correlations and weaken label privacy. We propose an optimized privacy-utility trade-off for label data by introducing a new definition for conditional feature-label differential privacy and enhancing the double randomized response (DRR) mechanism with a Rényi divergence optimizer. Additionally, a dynamic perturbation factor function improves usability across various tasks. Experimental results show that our method enhances resistance to label inference attacks by over 20% while maintaining similar accuracy, offering stronger privacy with a reduced privacy overhead and achieving an optimal balance during LLMs fine-tuning and inference.

Federated learning (FL) has become increasingly popular as a privacy-preserving, distributed training approach. However, recent studies have demonstrated that sensitive information can still be inferred from FL frameworks through certain attacks. Local differential privacy (LDP) provides privacy guarantees for FL and can help reduce potential privacy breaches. Despite this, current LDP-FL frameworks suffer from utility loss because the noise injected is directly proportional to the parameter dimension. To minimize noise, recent studies have employed private selection of the Top-k dimensions of the gradient vector. Yet, these methods are limited by their perturbations to individual data points and fail to address the dimensional dependency issue inherent in LDP. To overcome this challenge, we propose a novel joint perturbation method for sparse gradients, which is divided into two stages: perturbation of the gradient index vector and assignment of gradient values. Our method perturbs high-dimensional sparse gradient vectors as a whole, rather than individual gradients, thus reducing the noise injected into model gradients while preserving the necessary level of privacy. Moreover, to counteract the impact of the randomness introduced by perturbation on model performance, we integrate the Central Limit Theorem (CLT) into the gradient aggregation process, which we call CLT-Agg. We have validated our framework using public datasets, and our findings show a significant improvement over state-of-the-art methods. Extensive experiments have confirmed the effectiveness and efficiency of our proposed framework.

Federated learning enables participants to train on their local dataset, which solves the privacy preservation problem to some extent. However, attackers can still infer participants' private information from the uploaded parameters. Therefore, adding differential privacy further protects privacy. The key to differential privacy techniques is gradient clipping and gradient noise addition. In gradient clipping, a hyperparameter clipping threshold is introduced. Different combinations of clipping thresholds and learning rates lead to significant variations in accuracy, resulting in increased computational costs when searching for the optimal combination. In gradient noise addition, the gradient gradually decreases with training iterations. Adding fixed noise significantly affects the later stages of the model, leading to a decline in accuracy. In order to solve the above problems, this paper proposes An Adaptive Improvement Approach for Differential Privacy Federated Learning (AIDPFL), specifically (i) adjusting the clipping formula to combine the learning rate and the clipping thresholds under the premise of preserving the gradient information, only the learning rate needs to be adjusted. The clipping threshold size is adjusted in each round. (ii) Dynamically adjust the noise scale according to the gradient change, realize the dynamic decay rate to adjust the noise scale, and ensure that the privacy budget is reasonably allocated in the training process. Our method has higher usability and accuracy than the current primary adaptive differential privacy methods.

In federated learning, clients train models locally and upload only the model updates, while the server aggregates these updates through weighted averaging, effectively addressing the issues of data silos and privacy protection. However, in practical applications, some malicious clients may upload arbitrary model updates to the server, thereby hindering the convergence of the global model and ultimately leading to a decline in model performance. To address this problem, several Byzantine-robust aggregation schemes have been proposed. However, their effectiveness is significantly reduced in scenarios with data heterogeneity, and some robust schemes rely on prior knowledge of the number of malicious clients or require additional auxiliary datasets. Therefore, we propose a new Byzantine-robust aggregation scheme to mitigate the above issues. Specifically, we first use the median absolute deviation to identify anomalous gradient magnitudes and clip them to the median of the L2-norms, thereby avoiding excessively large gradient magnitudes that could alter the aggregation direction or lead to overly large updates. Next, we calculate the cosine distance between clients and apply the DBSCAN clustering method to group the updates, selecting the cluster with the largest number of elements as benign updates for aggregation while excluding malicious updates. Finally, an adaptive optimizer is employed on the server side to further mitigate data heterogeneity. Extensive experiments demonstrate that our method achieves superior performance under various Byzantine attacks in scenarios with data heterogeneity.

While prompt engineering is crucial for leveraging large language models (LLMs), existing optimization methods struggle to balance exploration of the vast prompt space with exploitation of high-quality candidates. To address this imbalance, we propose GRASP-Feedback Prompt (GFPrompt), a novel framework that integrates a dynamic grouping strategy with a teacher-model feedback mechanism. GFPrompt partitions prompts based on their performance: low-scoring prompts are sent to a global exploration operator to ensure diversity, while high-scoring ones undergo local refinement to enhance quality. A teacher model further provides semantic feedback to accelerate convergence. Experiments on six NLP tasks demonstrate that GFPrompt significantly outperforms standard baselines, expert-designed prompts, and state-of-the-art automated methods by up to 10.6%, 6.6%, and 1.4%, respectively. Furthermore, it reduces token consumption by up to 10.3% compared to leading evolutionary approaches under the same iteration count. Our results validate GFPrompt’s effectiveness in achieving superior prompt quality and computational efficiency.

With the growing importance of software supply chain security, automated identification of code reuse and dependency relationships among open-source components has become a critical task for building high-quality Software Bills of Materials (SBOMs). Existing code clone analysis methods such as CNEPS assist in component attribution and dependency extraction for C/C++ projects by utilizing function-level clone features. However, they exhibit significant shortcomings in complex multi-candidate component scenarios: their reliance on a single metric (clone function count) to determine the dominant component, combined with the lack of quantitative modeling for clone distribution uncertainty and component credibility, results in unstable dominant component selection and incorrect dependency edges.To address these limitations, we propose ConCloneDep, a method that preserves CNEPS’s module construction mechanism while introducing three key innovations: (1)a path-distance-first preliminary attribution strategy,(2)clone distribution entropy analysis, and (3)dominant component confidence modeling, thereby enabling quantitative evaluation of component attribution stability. For modules exhibiting high uncertainty, the method innovatively incorporates a neutral module mechanism to effectively mitigate the propagation of misjudged dependency edges. Experiments on 50 real-world open-source projects show that ConCloneDep achieves 85.2% dependency identification accuracy (versus CNEPS’s 79.1%) while maintaining 86.9% recall. It reduces incorrect dependency edges by 42.7% and improves average dependency edge confidence by 11%, significantly enhancing the reliability and accuracy of dependency analysis.

Differential privacy has become increasingly popular in private machine learning applications due to its provable ability to limit information leakage. However, there are often vulnerabilities in the practical implementation of differentially private algorithms, making it necessary to have effective tools for evaluating them before deployment. Unfortunately, the current state of the art classifier-based evaluation tools for differential privacy are still weakly distinguishable and need to be improved. In this paper, we propose a DP-Discriminator to automatically detect the ξ-differential distinguishability (ξ-DD) for specific algorithms, which is able to efficiently discover violations of differential privacy. Specially, we give a new attack definition of ξ-DD, based on a mathematical observation, which conduce to find a larger ξ. In addition, the proposed DP-Discriminator learns the overall distribution of features across samples depending on the ability of the generating adversarial network to capture the latent features. Benefiting from powerful classifiers, DP-Discriminator is able to automatically and accurately evaluate differential privacy with minimal time consumption. The experimental results demonstrate the effectiveness of the proposed method in estimating the ξ-DD for various practical randomized algorithms. For example, the latest work that detects the ξ-DD of the algorithm RAPPOR(0.4-DP) is 0.301, whereas our tool detects ξ-DD=0.369, with an error that is one order of magnitude lower.

Federated learning addresses the issues of data silos and privacy to some extent by training models locally on client devices, only uploading model parameters, and aggregating them on the central server. However, attackers can still infer private data information from the uploaded parameters. To solve this issue, differential privacy technology is introduced. The key to differential privacy lies in gradient clipping and noise addition. However, the traditional gradient clipping method often faces the gradient distortion issue and will be ineffective if the noise is large enough. Regarding noise addition, commonly used fixed or fixed decay rate noise scale settings often overlook the characteristics of gradients during training, which might lead to adding improper noise to gradients. To address these issues, we propose an adaptive optimization method for differentially private federated learning (AODPFL). Specifically, we adopt a strategy of clipping the gradient by grouping, effectively reducing gradient distortion. We design an adaptive clipping threshold based on the gradient changes during training to improve model accuracy under large noise conditions. We design a noise scale decay method with a dynamic decay rate to allocate the privacy budget more reasonably and inject appropriate noise into gradients. Experimental results show that compared to other gradient clipping and noise addition methods, our method achieves higher accuracy under the same privacy budget.

Textural Vulnerability Descriptions(TVD) refers to the natural language description of a vulnerability in databases like National Vulnerability Database(NVD) and China National Vulnerability Database(CNVD), which typically provides a concise summary of critical vulnerability details. To facilitate the understanding of domain-specific terms in TVD and the accurate extraction of information, we have introduced a bilingual domain-adaptation pre-trained model called BTVD-BERT, aimed at enhancing the model's capability to process and understand vulnerability descriptions in both Chinese and English. We explore three issues, the first being the impact of catastrophic forgetting on the model. Second, how should the dataset be proportioned to best enhance the model's generalization capabilities across both Chinese and English. Third, the addition of extra task-related metric information to the original dataset to construct a higher quality dataset, and whether training with this high-quality dataset can further improve model performance. We attempted to study the issues from a data engineering perspective and conducted numerous ablation experiments to find answers to these three questions. The experimental results indicate that catastrophic forgetting adversely affects the model, causing it to forget previously acquired knowledge while better retaining more recently obtained information. By employing a training approach using a mix of Chinese and English data, we were able to mitigate the impact of catastrophic forgetting on the model to some extent. Optimizing data proportions and improving data quality can effectively enhance the overall performance of the model. This study not only enhances the identification and analysis of security vulnerabilities but also offers new perspectives and empirical support for the research of multilingual domain-adaptive models.

Federated learning has become the leading paradigm for privacy-preserving distributed learning, as it only requires the upload of model gradients, not private data. However, recent studies have shown that these exchanged gradients can still lead to privacy leakage; for instance, attackers can recover private images through optimization-based gradient inversion attacks. In these attacks, pixel values in a dummy image are iteratively updated to approximate those of the private image. Yet, such optimization-based attacks often face challenges due to the large search space of the optimization task and struggle to accurately recover all labels in a batch. To tackle these, we propose the Optimization-based Feature-Level Gradient Inversion (OFLGI) attack, which focuses on optimizing image features rather than pixel values. This approach reduces the search space of the optimization task and enhances the quality of the reconstructed images. Our method formulates an optimization task to convert random noise features into natural image features, matching gradients while regularizing image fidelity. Notably, we are the first to utilize image features for gradient inversion attacks. Additionally, we introduce a novel batch label reconstruction algorithm that leverages the gradients of the biases in the final fully connected layer, surpassing existing methods even in the presence of duplicate labels. Furthermore, OFLGI demonstrates superior resistance to multiple gradient defense strategies, being able to recover high-quality private images even from degraded gradients. Extensive experiments have demonstrated the superiority of OFLGI over state-of-the-art gradient inversion techniques.

Despite the increasing participation of Chinese immigrant professionals in Australian workplace, they are still underrepresented in senior leadership positions and their perspectives have been overlooked in management and leadership research. Drawing on the literature on acculturation and leadership, this study explores the acculturation experiences of Chinese immigrant professionals (CIPs) and in turn their leadership perceptions, relative to a comparison group of Chinese Professionals (CPs) in China. We found that CIPs’ acculturation experiences influence their perceptions of ethics and respect for authority, but not their preference for participative decision making. Our study highlights the dynamic relationship between acculturation and key leadership issues from a follower’s perspective for immigrant professionals with a Chinese background. It extends current understanding of the cognitive outcomes of acculturation and has strong implications for cross-cultural leadership competency training, talent management and diversity and inclusion of minority workers.

Drawing on the conservation of resources (COR) theory, the authors investigate whether and how social support impacts the emotional labor of live streamers through professional identity. They also explore the boundary conditions by focusing on the moderation effect of emotional intelligence. Based on a sample of 331 live streamers in the content industry, the results show that social support weakens (enhances) live streamers' surface acting (deep acting) by enhancing their professional identity. Emotional intelligence significantly moderates the professional identity-emotional labor relationship. In addition, they find that emotional intelligence strengthens the negative indirect effect of social support on surface acting through professional identity but weakens the positive indirect effect of social support and deep acting through professional identity. They also discuss theoretical contribution in emotional labor literature and practical implications for live commerce.

Power supply for Internet of Things (IoT) devices is one of the bottlenecks in IoT development. To provide perpetual power supply for IoT devices, resonant beam charging (RBC) is a promising safe, long-range, and high-power wireless power transfer solution. How long distance RBC can reach and how much power RBC can transfer? In this paper, we analyze the consistent and steady operational conditions of the RBC system, which determine the maximum power transmission distance. Then, we study the power transmission efficiency within the operational distance, which determines the deliverable power through the RBC energy transmission channel. Based on the theoretical model of the wireless energy transmission channel, we establish a testbed. According to the experimental measurement, we validate our theoretical model. The experiments verify that the output electrical power at the RBC receiver can be up to 2 W. The maximum energy transmission distance is 2.6 m. Both the experimental and theoretical performance of the RBC system are evaluated in terms of the transmission distance, the transmission efficiency, and the output electrical power. Our theoretical model and experimental testbed lead to the guidelines for the RBC system design and implementation in practice.

In a research community, the provenance sharing of scientific workflows can enhance distributed research cooperation, experiment reproducibility verification and experiment repeatedly doing. Considering that scientists in such a community are often in a loose relation and distributed geographically, traditional centralized provenance sharing architectures have shown their disadvantages in poor trustworthiness, reliabilities and efficiency. Additionally, they are also difficult to protect the rights and interests of data providers. All these have been largely hindering the willings of distributed scientists to share their workflow provenance. Considering the big advantages of blockchain in decentralization, trustworthiness and high reliability, an approach to sharing scientific workflow provenance based on blockchain in a research community is proposed. To make the approach more practical, provenance is handled on-chain and original data is delivered off-chain. A kind of block structure to support efficient provenance storing and retrieving is designed, and an algorithm for scientists to search workflow segments from provenance as well as an algorithm for experiments backtracking are provided to enhance the experiment result sharing, save computing resource and time cost by avoiding repeated experiments as far as possible. Analyses show that the approach is efficient and effective.