Sequential recommendation has been widely studied for modeling users' dynamic preferences, and recent methods further incorporate social relationships to capture social influence. However, most social sequential recommendation models mainly rely on observed social links and sequence co-occurrence, while overlooking sparse and biased social graphs as well as latent item relationships. To address these limitations, we propose SRSI, an explicit–implicit synergistic social sequential recommendation framework. Instead of replacing explicit social relations, SRSI complements them with two implicit graphs. The implicit user graph is derived from interest similarity to alleviate missing or weak social links, while the implicit item graph is mined from frequent itemsets to capture cross-item purchasing depen dencies. These two graphs are injected into different encoders as structural inductive biases: the user graph expands social message propagation, and the item graph guides sequential item representation learning. Experiments on four datasets show that SRSI achieves strong and competitive Top-N recommendation performance. It obtains the best overall results on Epinions, Yelp, and LibraryThing, and remains competitive on Ciao. Further analysis shows that the effectiveness of implicit graph injection depends on graph quality, coverage, and noise. Ablation studies demonstrate that the two implicit graphs generally provide complementary signals for social sequential recommendation.
Security-critical values in container runtimes, services, and Kubernetes controllers often lose provenance when reduced to ordinary strings, paths, or JSON objects. This creates two distinct risks: application-local trust erasure, where insufficiently promoted values reach sensitive sinks, and cross plane semantic inconsistency, where admission and application artifacts fail to preserve the same predicate. We present T3C, a typed trust-contract framework that makes trust transitions explicit and co-generates application enforcement and admission policy for manifest-visible predicates. T3C provides boundary aware validation, sink-specific wrappers, and optional identity, freshness, and replay constraints, while reporting backend lowerings as exact, conservative, delegated, or unsupported. We implement T3C in a patched youki runtime and a SPIRE-backed order service. In youki, nine mount, masked-path, proc-write, and AppArmor sinks replace raw PathBuf/String inputs with typed wrappers, preventing unpromoted values from reaching privileged host operations at modest cost. In the service prototype, stale, replayed, and misbound requests are rejected, while read heavy paths remain sub-millisecond. Compared with schema derived validation and native typed wrappers, T3C adds little marginal runtime cost once equivalent identity and sink checks are enabled, reduces independently maintained enforcement artifacts, and exhibits zero observed application–policy disagreement across the controlled dual-plane revision corpus.
Exponential growth of network traffic has imposed tremendous pressure on the backhaul network; meanwhile, emerging applications (e.g., virtual reality) require extremely low latency. These challenges have exhausted the capability of the traditional cloud-based service framework. Edge caching is a potential way to address these issues by caching content at the base station (BS),11 In our system, the edge server is deployed on the base station. Unless stated otherwise, we use the terms base station and edge server interchangeably. which is closer to the terminal user, reducing service latency while alleviating backhaul link pressure. However, optimal content placement and cache management are critical for efficient edge caching. In this paper, we propose a deep reinforcement learning based edge caching scheme in a collaborative BS environment. Firstly, a BS recommendation algorithm is devised to optimize BS collaboration, which factors in both the distance between BSs and content similarity. Subsequently, we reformulate a cache replacement policy optimization problem as a partially observable Markov decision process and propose a multi-agent soft actor-critic algorithm (MASAC) for tackling the reconstructed problem. Furthermore, to fully utilize the collaboration information, we revamp the architecture of the actor network to extract more features of different content types and exploit a mask layer to filter out illegal content. Extensive simulation results demonstrate that the proposed MASAC algorithm significantly outperforms other existing caching schemes in reducing service latency and enhancing the forwarding request hit rate.
Grid-based location obfuscation is a widely used technique for protecting location privacy. In spatial crowdsourcing, several solutions combine this approach with cryptographic mechanisms to enable secure task assignment. However, they face a trade-off, incurring either high computational overhead or high storage overhead. To address this limitation, we propose a novel Grid-based Privacy-Preserving Scheme (GPPS). Our GPPS operates under a coarse-grained grid setting, motivated by the observation that larger grid cells provide stronger trajectory cloaking while reducing storage overhead. Concretely, we design a dynamic grid partitioning method based on online-worker density and employ a tree structure to represent hierarchical relationships among grid cells. Building upon this structure and an inner-product functional encryption algorithm, we develop a secure and accurate task assignment approach. Furthermore, a greedy strategy is adopted to improve assignment efficiency. We theoretically analyze the security of GPPS and evaluate its performance on two real-world datasets. Experimental results show that GPPS achieves a favorable trade-off between efficiency and storage overhead, making it a practical and scalable solution for large-scale spatial crowdsourcing services.
