Abstract

Internet of Things (IoT) devices such as smart phones and wireless sensors have proliferated in smart cities over the past few years. Various applications, including augmented reality, autonomous driving and smart homes, are based on the effective use of data generated by IoT devices. However, two main issues constrain the development of IoT applications. The first issue originates from the data distribution, which is usually isolated and not easy to be centralized due, in large part, to privacy concerns. The second issue arises from the shortage of labeled data, as the labeling process is costly, time-consuming and often requires input from domain experts. More recently, federated semi-supervised learning has become a viable solution to mitigate these issues by collaboratively training a machine learning model using decentralized labeled and unlabeled data. It also brings extensibility and generalizability without privacy infringement compared to traditional centralized training. However, the federated learning process is quite challenging due to data heterogeneity among clients. The contributions in this thesis are two-fold. First, we present an adaptive federated semi-supervised learning framework, which seamlessly integrates adaptive optimizers on both server and client sides in an effort to promote system adaptability. Experiments and ablations studies conducted on four standard benchmark datasets demonstrate the effectiveness of our proposed approach in image classification, achieving superior performance over strong baseline methods.
The other contribution consists of designing a two-stage human activity recognition system, which also incorporates adaptive optimizers into both local and global training. Clients train a local autoencoder model with a learning rate adaptive to local gradients, while the central orchestration server updates the global autoencoder model by applying a gradient-based adaptive optimizer to the average of clients’ model updates. Our system leverages a large amount of unlabeled data on clients with the aim of achieving a higher classification accuracy. The key benefit of adaptive optimizers is their ability to improve local training, while stabilizing the global aggregation in a bid to guarantee a proper optimization. We demonstrate through experiments that the proposed framework is robust to non-independent and identically distributed data and yields a stable convergence rate in different settings.