Collaborative Machine Learning (CML) enables edge devices to train models jointly without sharing raw data, offering a promising solution for privacy-sensitive and latency-critical applications in areas such as mobile computing, healthcare, and industrial automation. Among the various CML frameworks, Federated Learning (FL) has become the most widely adopted paradigm. However, deploying FL in real-world edge environments raises fundamental trust-related challenges, each linked to one of the three core components of CML: local data, learning model, and collaboration mechanism. Specifically, at the data level, strong privacy guarantees are difficult to achieve due to the risk of gradient leakage. At the model level, machine unlearning becomes essential to remove outdated, erroneous, or sensitive data from trained models. At the collaboration level, robustness against adversarial or faulty clients is critical to ensure reliable training in fully open environments. These three aspects are inherently connected: if privacy risks are uncontrollable, users will be hesitant to contribute data; if the model lacks unlearning capabilities, the system will be unable to adapt to data deletions and long-term evolution; and if the collaborative mechanism lacks robustness, the training process will struggle to converge. Guided by this logic, this dissertation systematically addresses these challenges by developing trustworthy FL algorithms with formal theoretical guarantees on convergence, differential privacy, machine unlearning, and Byzantine robustness. First, to mitigate privacy risks in wireless edge learning, the dissertation explores FL over analog over-the-air aggregation channels, where both inherent wireless noise and participant-injected perturbations jointly contribute to privacy amplification. A privacy-aware OTA-FL algorithm is proposed that jointly leverages wireless noise and structured perturbation to achieve user-level local differential privacy, even under malicious server-side manipulation of channel state information (CSI). A bandwidth-adaptive gradient compression scheme is integrated to reduce communication overhead and accommodate limited bandwidth. Theoretical guarantees on privacy and convergence are established, and experimental results demonstrate robust protection against CSI attacks, achieving improved trade-offs between privacy, accuracy, and communication efficiency compared to state-of-the-art OTA-FL baselines. Second, to address the need for data forgetting, the dissertation introduces a provably exact federated unlearning framework. Exact unlearning is defined as the requirement that the distribution of the output model and all intermediate states of the unlearning algorithm be statistically indistinguishable from those of a retraining process conducted from scratch without the deleted data. By establishing a novel connection between federated unlearning and an optimal transport problem, the work identifies a sufficient condition for exact unlearning: the underlying FL algorithm must satisfy Total Variation (TV) stability. Guided by this insight, a TV-stable FL algorithm called FATS is proposed, based on local SGD with periodic averaging and structured sub-sampling. Efficient unlearning protocols are further developed for both sample-level and client-level deletion. Theoretically, FATS is proven to guarantee exact unlearning with favorable convergence rates. Empirical evaluations across multiple benchmarks confirm that FATS enables fast and thorough unlearning with significantly reduced computation and communication overhead compared to prior methods. Third, to ensure robustness against unreliable or malicious devices, the dissertation investigates FL under heavy-tailed gradient noise. It first considers scenarios where gradients follow coordinate-wise finite-variance heavy-tailed distributions, and proposes a local soft truncation-based gradient processing strategy that, when combined with appropriate global aggregation, achieves optimal Byzantine resilience in homogeneous data settings. The work is then extended to more realistic conditions involving infinite-variance noise and statistical heterogeneity. Two robust algorithms are developed using gradient and momentum clipping techniques. Additionally, a random projection-based approximate neighbor mixing method is introduced to reduce aggregation overhead in high-dimensional scenarios. On the theoretical side, the dissertation establishes the first high-probability convergence bounds under these relaxed assumptions. Empirically, the proposed algorithms demonstrate strong performance in adversarial and heterogeneous environments where previous approaches fail. Although the dissertation focuses on the FL framework, the underlying algorithmic principles and theoretical insights generalize to other CML paradigms such as split learning, swarm learning, and peer-to-peer learning. This work thus lays a cohesive foundation for trustworthy CML at the edge, supporting the secure and reliable deployment of intelligent edge systems.
Youming Tao (Thu,) studied this question.