Chain-of-Thought (CoT) prompting plays a critical role in enhancing the performance of large language models on complex tasks. However, CoT is vulnerable to attacks like BADCHAIN that exploit reasoning weaknesses. In this paper, we propose a novel approach, called Program Verification of Chain-of-Thought (PVCoT), which transforms CoT reasoning into a program verification task. By converting natural language logic into code, we enable the use of program verification tools to ensure the correctness and security of the reasoning process. If the generated program meets predefined properties, the reasoning chain is executed; otherwise, the problem is restated to derive the answer. Specifically, we leverage the SMT solver to validate whether a generated program satisfies predefined properties. Depending on whether the properties are satisfied or not, the reasoning chain is either directly executed or Problem Restatement, providing an effective defense mechanism. Experimental results shows that modeling CoT defenses as program verification tasks is an effective method to defend against existing CoT attack algorithms and significantly outperforms existing defense methods.

Label Propagation Algorithm (LPA) is a method of predicting node labels on graph structured data by utilizing the correlation relationships between nodes.It is widely used in scenarios such as community detection and risk transmission analysis.Current LPA is mainly executed on a single data source, and the data value of individual participants is usually limited.With the increasing demand of data applications in multiple participants, integrating multi-party data for label propagation is a promising research direction.In multi-party scenarios, each participant usually needs to avoid exposing their own private data to the other party.How to enable multiple participant to better execute LPA while protecting their own privacy data, and achieve effective integration and utilization of multi-party data value, is a problem that needs to be studied.This article proposes a privacy preserving cross-domain label propagation algorithm (PPCD-LPA), which uses a secure intersection protocol based on RSA+HASH algorithm to achieve secure matching of intersection nodes in the graph structured data of each participating party.A secure label aggregation algorithm based on semi-homomorphic encryption and a combination of random numbers is designed to calculate and update the labels of each node, thereby protecting the sensitive information of each party.This article tested and validated PPCD-LPA on multiple datasets, and the experimental results showed that PPCD-LPA can run securely and correctly among multiple parties, achieve the same model effect as plaintext cross-domain label propagation algorithm and has engineering practicality.

One important feature of Content-centric Networking (CCN) is to utilize built-in network caches to improve the transmission efficiency of content dissemination. In CCN, all nodes have the function of caching contents. Consequently, caches are ubiquitous and diverse. How to make full use of these caches to reduce the network load and the transmission delay has become a hot research topic in recent years. In order to speed up content distribution and improve network resource utilization, a cooperative caching strategy called Tracker-based Cache Utilization Strategy (TBCUS) is proposed in this paper. In TBCUS, an idea of partial centralized management of the caches is introduced into traditional CCN. Moreover, in order to alleviate the pressure that the incurred overhead imposes on network bandwidth, TBCUS divides the whole caching system into several sub-systems. Each of the sub-systems has a tracker server as the controller and some peers. The contents cached in a sub-system are transparent to each other through the coordination of the tracker server. Once a content request reaches to a sub-system, all the caches in the sub-system can be selected to respond to this request. While the traditional caching strategy in CCN is that only the caches on the delivery path can be used to respond to the content request. TBCUS improves the cache utilization and has higher cache hit ratio and lower transmission delay and network load compared with the traditional caching strategy.

Content-centric Networking (CCN) is a new Internet architecture, relies on in-network caching to achieve an efficient and reliable content distribution. However, the existing forwarding strategy cannot fully sense the location of cache, because it is rely on the ports of routing table. In this paper, we propose a Bidirectional Cache Locating Algorithm based on Ant Colony Optimization (BCLA) using bidirectional ants to diffuse and exploit multiple content copies. One direction of ant called Announcing Ant is responsible for advertising the information of cached content to neighbors in a limited area, the other side, including Probing Ant and Updating Ant, are responsible for seeking cached content for the routers outside the area. Our goal is to find the optimal cache, and making efficient utilization of available cache resources within a certain area. Extensive experiments have been carried to evaluate BCLA against other forwarding strategies. The simulation results prove that BCLA outperforms other strategies in the content delivery latency, cache hit ratio, as well as the path stretch.