Data Preparation (DP), which involves tasks such as data cleaning, imputation and integration, is a fundamental process in data-driven applications. Recently, Large Language Models (LLMs) fine-tuned for DP tasks, i.e., DP-LLMs, have achieved state-of-the-art performance. However, transferring DP-LLMs to novel datasets and tasks typically requires a substantial amount of labeled data, which is impractical in many real-world scenarios. To address this, we propose a knowledge augmentation framework for data preparation, dubbed KNOWTRANS. This framework allows DP-LLMs to be transferred to novel datasets and tasks with a few data points, significantly decreasing the dependence on extensive labeled data. KNOWTRANS comprises two components: Selective Knowledge Concentration and Automatic Knowledge Bridging. The first component re-uses knowledge from previously learned tasks, while the second automatically integrates additional knowledge from external sources. Extensive experiments on 13 datasets demonstrate the effectiveness of KNOWTRANS. KNOWTRANS boosts the performance of the state-of-the-art DP-LLM, Jellyfish-7B, by an average of 4.93%, enabling it to outperform both GPT-4 and GPT-4o.

Deep neural networks (DNNs) have driven breakthroughs in autonomous driving through end-to-end methods, utilizing their powerful learning capabilities to generate vehicle controls directly from sensor data. However, maximizing the satisfaction of DNN inference requirements under the constraints of limited computing and energy resources on the vehicle side has emerged as a critical challenge in Vehicular Edge Computing(VEC). To address these challenges, we propose the reinforcement learning-empowered task diversion scheduling algorithm named RTD. This algorithm intelligently offloads computationally intensive portions of the DNN to Roadside Units (RSUs) by taking into account factors such as the battery coefficient and the type of DNNs. Firstly, we utilize the FLOPs method to model the data flow structure and computational load distribution of the DNNs. Subsequently, we formulate the task offloading model as an optimization problem that jointly considers latency, energy consumption, and the remaining battery power of the vehicle. Finally, after simplifying the optimization problem using the diversion algorithm, we employ the SAC method to determine the optimal offloading strategy. Extensive experiments demonstrate that RTD significantly reduces overall task completion time, effectively handles time-sensitive tasks, properly protects low-battery vehicles, and adapts well to dynamic network environments.

The adoption of collaborative robots is rapidly advancing and enhancing efficiency. However, safety in human-machine interaction has long been a crucial concern in various fields, including manufacturing plants and product development. Existing standards, such as ISO 10218 and ISO/TS 15066, set force and pressure limits based on pain thresholds. Improving safety in collaborative tasks requires better pain assessment and monitoring contact conditions. Given individual differences in pain perception, refined safety standards should consider lighter, non-harmful contact and integrate psychological and social pain factors to enhance human-robot interaction safety. Creating a human body dummy for interaction simulation experiments is an effective experimental method. In this study, mechanical stimulus experiments were conducted simulating pinch environments on a human finger within a safe range. Using a pain assessment tool, the relationship between pain intensity, pain cycle, and contact force was analysed considering both sensory and emotional pain factors. Furthermore, a hand dummy equipped with phalange dummy modules with built-in flexible sensors was designed. The effectiveness of distinguishing different contact conditions was evaluated through a friction contact experiment.

Federated learning (FL) is highly susceptible to backdoor attacks, which cause the global model to produce incorrect prediction results for specific inputs. Existing defense methods primarily rely on clustering to exclude significantly deviated model updates from aggregation, or use differential privacy or neuron pruning to limit the impact of backdoors, which only are effective against specific attacks or significantly degrade the global model's usability. Moreover, these methods require the appropriate hyperparameters. To address the aforementioned limitations, we propose MPRA (Multi-Perspective Resilient Aggregation). First, we detect the differences between benign and malicious model updates from three different perspectives: the deviation angle between model updates, the magnitude of neuron updates, and the distance between model updates. Then, the model updates that fail to pass the detection will be regarded as latent malicious model updates and excluded from the aggregation. Extensive experiments on three different datasets show that the proposed MPRA can effectively tackle various backdoor attacks and minimize the negative impact on the global model's usability in the majority of scenarios without acquiring prior knowledge such as attack rounds, attack scale and attack methods.

Federated recommender systems (FedRecs) have received a lot of attention in recent years because of their excellent ability to protect privacy. But recent research has shown that federated recommender systems are vulnerable to poisoning attacks. Attackers can pollute training data or local models to influence the recommendation rankings of target items. To reduce the impact of poisoning attacks on federated recommender systems, we propose a robust federated recommendation framework based on target item detection called FedRecTID, which can effectively resist the impact of targeted attacks in federated recommender systems. FedRecTID can be divided into three steps, including the identification of suspicious items, vote for validation and removal of the effects of attack. Specifically, after receiving the gradients uploaded by the clients, the server calculates the suspicious scores of items by calculating the cosine similarity of the item embedding in the last two rounds. Based on the scores, the suspicious items can be determined by statistical anomaly detection methods. Then these items are sent to the selected users for verification. After receiving users’ vote, the server validates it based on the majority of the voting mechanism, and filters out the target item without updating it. Experimental results on two real datasets demonstrate that our proposed framework for FedRecs can effectively defend existing targeted attacks against FedRecs. And FedRecTID does not harm the recommendation performance.

Graph neural networks (GNNs) are widely utilized in recommender systems to model higher-order interactions between users and items due to their capacity to handle the intricate relational data. Federated Learning, known for its excellent privacy-preserving capabilities, is extensively applied in recommender systems. Nonetheless, existing GNN-based recommendation methods fail to fully leverage the integration of social networks and knowledge graphs into the recommendation process, neglecting the subtle differences between these data sources. Furthermore, current federated learning-based recommendation frameworks cannot balance efficiency and privacy protection, with adversaries potentially exploiting attacks to breach user privacy. To address these issues, we propose Federated Knowledge-enhanced Graph Attention Network for Social Recommendation (FGAT4SR) to achieve efficient recommendations. Specifically, we use graph attention networks to learn the rich knowledge and aggregate their most critical information. Then pseudo-item interactions and local differential privacy are incorporated into our model to ensure privacy protection. Extensive experiments on real datasets demonstrate that our proposed federated recommendation framework outperforms all baseline methods. Our approach enables the sophisticated integration of social relationships and knowledge graphs while preserving privacy.

Advances in the field of visual-language contrastive learning have made it possible for many downstream applications to be carried out efficiently and accurately by simply taking the dot product between image and text representations. One of the most representative approaches proposed recently known as CLIP has quickly garnered widespread adoption due to its effectiveness. CLIP is trained with an InfoNCE loss that takes into account both positive and negative samples to help learn a much more robust representation space. This paper however reveals that the common downstream practice of taking a dot product is only a zeroth-order approximation of the optimization goal, resulting in a loss of information during test-time. Intuitively, since the model has been optimized based on the InfoNCE loss, test-time procedures should ideally also be in alignment. The question lies in how one can retrieve any semblance of negative samples information during inference in a computationally efficient way. We propose Distribution Normalization (DN), where we approximate the mean representation of a batch of test samples and use such a mean to represent what would be analogous to negative samples in the InfoNCE loss. DN requires no retraining or fine-tuning and can be effortlessly applied during inference. Extensive experiments on a wide variety of downstream tasks exhibit a clear advantage of DN over the dot product on top of other existing test-time augmentation methods.