Recently, self-play fine-tuning (SPIN) has garnered widespread attention as it enables large language models (LLMs) to iteratively enhance their capabilities through simulated interactions with themselves, transforming a weak LLM into strong one. However, applying SPIN fine-tune text-to-SQL presents substantial challenges. Notably, existing frameworks lack clear signal feedback during the training process and fail adequately capture implicit schema-linking characteristics between natural...

In this paper, we propose a method, namely Goalie, to defend against the correlated value and sign encoding attacks used steal shared data from trusts. Existing methods prevent these by perturbing model parameters, gradients, or training while significantly degrading performance. To guarantee performance of benign models, Goalie detects malicious models stops their training. The key insight detection is that additional information in parameters through regularization terms changes parameter...

Anomaly detection for log sequences is a necessary task system intelligent operation and fault diagnosis. In sequence, adjacent logs have the property of local correlation, while long-distance remote dependencies. It helpful to fully mine these information during modeling improving performance anomaly detection. Meanwhile, there are some redundant or noise in which has no contribution detection, may even bring negative impact. The existing methods sequence do not take above problems into...

Searchable Encryption (SE) has been extensively examined by both academic and industry researchers. While many SE schemes show provable security, they usually expose some query information (e.g., search access patterns) to achieve high efficiency. However, several inference attacks have exploited such leakage, e.g., a recovery attack can convert opaque trapdoors their corresponding keywords based on prior knowledge. On the other hand, proposed require significant modification of existing...

The multi-label recognition of damaged waste bottles has important significance in environmental protection. However, most the previous methods are known for their poor performance, especially regards to bottle classification. In this paper, we propose use a serial attention frame (SAF) overcome mentioned drawback. proposed network architecture includes following three parts: residual learning block (RB), mixed (MAB), and self-attention (SAB). RB uses ResNet pretrain SAF extract more...

Searchable encryption (SE) schemes, such as those deployed for cyber-physical social systems, may be vulnerable to inference attacks. In attacks, attackers seek learn sensitive information about the queries and data stored on (cyber-physical social) systems. However, these attacks are often based strong (impractical) assumptions (e.g., complete knowledge of documents or known document injection) using access-pattern leakage. this paper, we first identify different leakage models profiles...

Large Language Model-based (LLM-based) Text-to-SQL methods have achieved important progress in generating SQL queries for real-world applications. When confronted with table content-aware questions scenarios, ambiguous data content keywords and non-existent database schema column names within the question leads to poor performance of existing methods. To solve this problem, we propose a novel approach towards Table Content-aware Self-Retrieval (TCSR-SQL). It leverages LLM's in-context...

Encryption will play a more important role as computation and storage are outsourced. Achieving good balance between strong security application functionality preservation becomes cutting-edge research problem. SE algorithms well studied, which delegate search capabilities to the cloud provider without decrypting documents. But this approach imposes extra constraints on API loses query expressiveness. This paper proposes CASB based framework for encrypted data sharing, builds index locally...

Browser-based cloud storage services are still broadly used in enterprises for online sharing and collaboration. However, sensitive information images or documents may be easily leaked outside trusted enterprise on-premises due to such services. Existing solutions prevent data leakage either limit many functionalities of applications difficult scaled various applications. In this paper, we propose CloudDLP, a transparent scalable approach automatically sanitize with browser-based CloudDLP is...

According to the statistics of BrightPlanet in 2012, information contained Deep Web is 400-500 times more than those Surface Web. Automatic data collection under has become one research hotspots domain web crawlers. During exploration automatic dynamic crawlers, we discover two inevitable situations which have not been processed properly by existing tools. One that some websites adopt CSS pseudo-class locate clickable elements, prevents technologies from simulating user actions. Another...

Confidential data is often encrypted before it uploaded to cloud servers. However, client-controlled encryption poses a major barrier towards the full functionalities of services. This paper presents SafeBox, new Cloud Access Security Broker (CASB)-based approach that protects sensitive information against attackers with control servers, and allows clients search share transparently. It addresses following challenges: First, SafeBox brings almost no loss for protecting in applications....

Swarm learning (SL) is an emerging promising decentralized machine paradigm and has achieved high performance in clinical applications. SL solves the problem of a central structure federated by combining edge computing blockchain-based peer-to-peer network. While there are results assumption independent identically distributed (IID) data across participants, suffers from degradation as degree non-IID increases. To address this problem, we propose generative augmentation framework swarm...

The primary business challenge for the customers to use outsourced computation and storage is loss of data control security. So encryption will become a commodity in near future. There big diffusion with above scenario: take advantage current application’s full functionalities at same time ensuring their sensitive remains protected under customers’ control. Prior works have achieved effective progress towards satisfying both sides. But there are still some technical challenges, such as...

Federated learning (FL) allows decentralized medical institutions to collaboratively learn a shared global model without breaching data privacy. However, in the context of image segmentation, distributions across centers may vary lot due diverse imaging protocols, vendors and partial annotation, which usually hampers optimization convergence performance FL. In this paper, we propose novel approach called federated knowledge augmentation (FedKA) address non-IID (non-independent identically...

Abstract Real‐world data always exhibit an imbalanced and long‐tailed distribution, which leads to poor performance for neural network‐based classification. Existing methods mainly tackle this problem by reweighting the loss function or rebalancing classifier. However, one crucial aspect overlooked previous research studies is feature space caused angle distribution. In paper, authors shed light on significance of distribution in achieving a balanced space, essential improving model under...