With the increasingly stringent data privacy protection regulations and the decentralisation of computing resources, federated learning has gradually become a current research hotspot. Currently federated learning can effectively solve the problem of data silos, but there are still several notable challenges in the current federated learning in specific application practices, including the lack of efficient and secure incentive mechanisms, the increasingly obvious drawbacks of the traditional centralised management model, and the training security problems caused by malicious users. This paper is dedicated to solving these problems by proposing a secure federated learning mechanism based on a dual incentive strategy of data and arithmetic resources, aiming to improve the performance and security of the system through innovative algorithms and framework design.

This paper proposes a dual metric mechanism based on data resources and arithmetic resources. Contribution measurement is the antecedent problem of designing the federated learning incentive mechanism, this paper improves the data resources and arithmetic resources of the federated training client to be evaluated from multiple dimensions, and selects multiple indicators for specific measurements, in which the innovative introduction of task relevance indicators can more accurately match the client with the learning task, which further improves the resource utilisation rate and the efficiency of federated training.

As the data trading market continues to expand, the issues of privacy information leakage, data leakage, and data quality in data trading are becoming increasingly prominent, and to a certain extent, constraining the development of the data trading market. Blockchain, as a distributed ledger technology, with its characteristics of decentralization, traceability, and tamper resistance, can effectively avoid problems such as single point of failure and data tampering in data trading. This paper addresses the issues of privacy information leakage, data quality, and transaction fairness in data trading. Firstly, this paper introduces a decision tree classification model to classify the source data, and removes data that does not meet the transaction requirements before data trading. During the secure comparison phase, a lightweight ciphertext comparison algorithm is designed so that data owners can obtain classification results without revealing plaintext data to the decision tree classification model.

During the data trading phase, this paper conducts transactions on a per-data fragment basis, granting data buyers trial rights. Subsequently, upon data buyers obtaining transaction data, consistency verification is conducted on the transaction data to prevent data owners from adding garbage data to the transaction data before data trading, thereby avoiding dishonest trading behavior by data owners. Finally, this paper designs a decentralized data trading model, DCTM (Data Classification Transaction Model), based on smart contract technology. This model addresses issues such as data quality, single point of failure, privacy leakage, and data resale in data trading, ensuring transaction fairness and security

Traditional evolutionary algorithms are often used to solve single-task optimization problems. With the study of evolutionary algorithms, it is found that most optimization tasks often have potential correlation, which indicates that the knowledge obtained in the evolution process of one optimization task can be used to optimize another task to further optimize the performance of the target task. With the rapid development of machine learning technology, the idea of making use of commonalities or differences among multiple tasks for efficient learning has been widely studied in the field of multi-task learning.In order to strengthen the positive transfer of knowledge, we can consider using strong correlation prior knowledge to construct help tasks and optimize the original tasks together. In this paper, a method of constructing help tasks is proposed, which is based on the original problem and constructed by multi-objective decomposition of sub-problem groups. These sub-problem sets are closely related to the original problem and belong to the strongly correlated tasks. Therefore, the help tasks constructed by this method can promote the forward knowledge transfer of multi-task optimization theoretically. The experimental results show that the efficiency of the proposed algorithm is improved significantly in task optimization.

In recent years, the role of monitoring systems in upholding public order and ensuring public safety has expanded. Video monitoring in locations such as train stations, schools, and hospitals necessitates the automatic identification of hostile and suspicious behaviours to avert any potential harm to society, the economy, and the environment. The optimal utilization of computerized violence detection by law enforcement is crucial. An all-encompassing methodology is necessary to identify instances of violence and the use of weapons in closed circuit television (CCTV) recordings. This work presents the Smart-City CCTV Violence Detection (SCVD) dataset, which is used for detecting occurrences of violence in surveillance films, including both cases including weapons and those without weapons. The proposed system utilizes a model based on deep learning approach to detect violence in videos. The model uses a comprehensive set of characteristics to describe violence occurrences in the dataset. The model was created using MATLAB software, utilizing image processing and machine and deep learning toolkit. It achieved an overall accuracy rate of 96.4%. The proposed endeavour has practical applications in the industry and is advantageous to society in terms of security

In recent years, with the emergence of vulnerabilities such as "Spectre" and "Meltdown", efficient testing of processor security has become an important issue for research. Fuzz testing has also been applied in processor security testing in recent years. However, genetic algorithms exhibit a certain degree of blindness when performing fine-grained mutation, which can lead to the generation of test cases that trigger the same type of vulnerabilities, thereby failing to cover more types of vulnerabilities in a short period of time. To address the above issues, a guided processor security fuzzing scheme based on DQN reinforcement learning is proposed. This scheme constructs a reward function by associating the state values of test cases with the weights corresponding to the triggered vulnerability types. It employs reinforcement learning to guide the generation of test cases with targeted and directional mutations, thereby quickly triggering different types of vulnerabilities. Currently, there is no similar fuzz testing system in the industry. Experiments conducted on the  ARMv8 platform demonstrate that compared to traditional approaches using genetic algorithms as feedback, this scheme significantly reduces the required time while covering an equivalent number of branch prediction vulnerabilities. The efficiency of vulnerability detection is improved by 1.29 times.

The purpose of single-image dehazing technology is to recover a clean image from hazy inputs, reduce the impact of weather on image quality, improve the visual effect of the image, and are essential for subsequent advanced tasks, such as image recognition and scene understanding. Due to the rapid development of  computer technology and deep learning for image processing, researchers have the ability to dehaze foggy images, especially in recent years, the innovation of image dehazing technology is used to support more accurate advanced tasks. Based on the perspectives of end-to-end learning and physical modeling, this paper collates the single-image dehazing papers of the top computer vision conference in the past three years from 2018 to 2020, and discusses the two types of articles and experimental results in depth. At the end of this paper, the problems existing in the current image dehazing technology and the direction of improvement are analyzed, and the development trend of single-image dehazing algorithm is prospected.