In recent years, the field of distributed deep learning within the Internet of Things (IoT) or the edge has experienced exponential growth. Federated meta-learning has emerged as a significant advancement, enabling collaborative learning among source nodes to establish a global model initialization. This approach allows for optimal performance while necessitating minimal data samples for updating model parameters at the target node. Federated meta-learning has gained increased attention due to its capacity to provide real-time edge intelligence. However, a critical aspect that remains inadequately explored is the recovery of interim meta knowledge’s failure, which constitutes a pivotal key for adapting to new tasks. In this paper, we introduce FMRec, a novel platform designed to offer a fast and flexible recovery mechanism for failed interim meta knowledge in various federated meta-learning scenarios. FMRec serves as a complementary system compatible with different types of federated models and is adaptable to diverse tasks. We present a demonstration of its design and assess its efficiency and reliability through real-world applications.

Today, large-scale cloud organizations are deploying datacenters and “edge” clusters globally to provide low-latency access to services. Running stream applications across geo-distributed sites are emerging as a daily requirement. However, existing efforts have dominantly centered aroundstateless stream processing, leaving another urgent trend-stateful stream processing-much less explored. A driving need is to store and update states during processing, and most importantly, successfully recover large distributed states when faults and failures happen. Existing studies exhibit major limitations including: (1) they mostly inherit MapReduce's “single master/many workers” architecture, where the central master can easily become ascalability bottleneck; (2) they offer state recovery mainly through three approaches: replication recovery, checkpointing recovery, and DStream-based lineage recovery, which are either slow, resource-expensive or failing to handle multiple failures; and (3) they are not adaptive to heterogeneous hardware settings. We present A-FP4S, a novel adaptive fragments-based parallel state recovery mechanism for stream processing systems. A-FP4S organizes stream operators into a distributed hash table based peer-to-peer overlay and divides each node's local state into many fragments. These fragments are periodically stored in node's multiple neighbors, ensuring different sets of available fragments can reconstruct failed states in parallel. This mechanism is extremely scalable to the lost state, significantly reduces failure recovery time, and can tolerate multiple node failures. A-FP4S is adaptive to heterogeneous hardware settings by automatic parameter tuning over phases. Compared to Apache Storm, A-FP4S achieves 31.8% to 50.5% reduction in recovery latency. Large-scale experiments using real-world datasets demonstrate A-FP4S's attractive scalability and adaptivity properties.

Music recommendations are normally based on a users prior artist and genre preference, or based on a similar users preference. This may result in recommendations for users being limited to subsections of artists and sub-genres, rather than offering an exploration of different genres that are similar to the user's vocal preferences. In this work, we propose a federated learning-based approach that scalably tunes music clusters to accurately describe users' preferences in a particular genre. It uses deep clustering on frequency sets and mel-spectrograms of songs. It can improve song recommendations based on the user's musical tastes.

Streaming computations are by nature long-running. They run in highly dynamic distributed environments where many stream operators may leave or fail at the same time. Most of them are stateful, in which stream operators need to store and maintain large-sized state in memory, resulting in expensive time and space costs to recover them. The state-of-the-art stream processing systems offer failure recovery mainly through three approaches: replication recovery, checkpointing recovery, and DStream-based lineage recovery, which are either slow, resource-expensive or fail to handle many simultaneous failures.We present FP4S, a novel fragment-based parallel state recovery mechanism that can handle many simultaneous failures for a large number of concurrently running stream applications. The novelty of FP4S is that we organize all the application's operators into a distributed hash table (DHT) based consistent ring to associate each operator with a unique set of neighbors. Then we divide each operator's in-memory state into many fragments and periodically save them in each node's neighbors, ensuring that different sets of available fragments can reconstruct lost state in parallel. This approach makes this failure recovery mechanism extremely scalable, and allows it to tolerate many simultaneous operator failures. We apply FP4S on Apache Storm and evaluate it using large-scale real-world experiments, which demonstrate its scalability, efficiency, and fast failure recovery features. When compared to the state-of-the-art solutions (Apache Storm), FP4S reduces 37.8% latency of state recovery and saves more than half of the hardware costs. It can scale to many simultaneous failures and successfully recover the states when up to 66.6% of states fail or get lost.

Modern stream processing applications need to store and update state along with their processing, and process live data streams in a timely fashion from massive and geo-distributed data sets. Since they run in a dynamic distributed environment and their workloads may change in unexpected ways, multiple stream operators can fail at the same time, causing severe state loss. However, the state-of-the-art stream processing systems are mainly designed for low-latency intra-datacenter settings and do not scale well for running stream applications that contain large distributed states, suffering a significantly centralized bottleneck and high latency to recover state. They offer failure recovery mainly through three approaches: replication recovery, checkpointing recovery, and DStream-based lineage recovery, which are either slow, resource-expensive or fail to handle multiple simultaneous failures.

There has been a dramatic increase in the popularity of Container as a Service (CaaS) clouds. The CaaS multi-tier applications could be optimized by using network topology, link or server load knowledge to choose the best endpoints to run in CaaS cloud. However, it is difficult to apply those optimizations to the public datacenter shared by multi-tenants. This is because of the opacity between the tenants and the datacenter providers: Providers have no insight into tenant's container workloads and dependencies, while tenants have no clue about the underlying network topology, link, and load. As a result, containers might be booted at wrong physical nodes that lead to performance degradation due to bi-section bandwidth bottleneck or co-located container interference. We propose 'DocMan', a toolset that adopts a black-box approach to discover container ensembles and collect information about intra-ensemble container interactions. It uses a combination of techniques such as distance identification and hierarchical clustering. The experimental results demonstrate that DocMan enables optimized containers placement to reduce the stress on bi-section bandwidth of the datacenter's network. The method can detect container ensembles at low cost and with 92% accuracy and significantly improve performance for multi-tier applications under the best of circumstances.

Web-based social networks enable new community-based opportunities for participants to engage, share their thoughts, and interact with each other. Theses related activities such as searching and advertising are threatened by spammers, content polluters, and malware disseminators. We propose a scalable spam detection system, termed Oases, for uncovering social spam in social networks using an online and scalable approach. The novelty of our design lies in two key components: (1) a decentralized DHT-based tree overlay deployment for harvesting and uncovering deceptive spam from social communities; and (2) a progressive aggregation tree for aggregating the properties of these spam posts for creating new spam classifiers to actively filter out new spam. We design and implement the prototype of Oases and discuss the design considerations of the proposed approach. Our large-scale experiments using real-world Twitter data demonstrate scalability, attractive load-balancing, and graceful efficiency in online spam detection for social networks.

Disinformation in social networks has been a worldwide problem. Social users are surrounded by a huge volume of malicious links, biased comments, fake reviews, or fraudulent advertisements, etc. Traditional spam detection approaches propose a variety of statistical feature-based models to filter out social spam from a historical dataset. However, they omit the real word situation of social data, that is, social spam is fast changing with new topics or events. Therefore, traditional approaches cannot effectively achieve online detection of the "drifting" social spam with a fixed statistic feature set. In this paper, we present Sifter, a system which can detect online social spam in a scalable manner without the labor-intensive feature engineering. The Sifter system is two-fold: (1) a decentralized DHT-based overlay deployment for harnessing the group characteristics of social spam activities within a specific topic/event; (2) a social spam processing with the support of Recurrent Neural Network (RNN) to get rid of the traditional manual feature engineering. Results show that Sifter achieves graceful spam detection performances with the minimal size of data and good balance in group management.