Vertical split learning (VSL) has recently emerged as a novel privacy-preserving paradigm by partitioning a model between multiple clients and a server. Despite its practical utility, recent research has revealed its vulnerability to backdoor attacks, where malicious attackers inject poisoned samples embedded with crafted triggers into the training data. In this paper, we present a stealthyTargetedPoisoningAttack within the context of VSL, termed TPA-VSL, which directly manipulates the embedding model without introducing any obvious trigger patterns. The crux of TPA-VSL is to map the embedding vector of the targeted sample to the attacker-desired class, adversely affecting the targeted sample's prediction. To achieve this, TPA-VSL features two novel components. The first component leverages the conditional generation capability of the state-of-the-art generative models — diffusion models, and uniquely guides them with an integrated multimodal encoder-decoder for informative training data generation. This approach allows us to mimic the target model and obtain the mappings of the targeted sample in the embedding space. The second component effectively poisons the embedding model by aligning the mappings of the targeted samples with those of the attacker-desired class. Experimental results demonstrate that TPA-VSL can achieve a 30% higher attack success rate on average compared to baseline attacks.

With the prevailing Mixture-of-Experts (MoE) architecture pushing the performance of Large Language Models (LLMs) to new limits, fine-tuning MoE models presents a significant challenge due to their tremendous number of parameters and sparsely activated network structures. While expert parallelism has been proposed to train large-scale MoE models by distributing expert layers among multiple devices, it fails to exploit the unique communication patterns in fine-tuning pre-trained MoE models. In this paper, we demonstrate that expert layers are not uniformly accessed, but exhibit a stable locality, with some experts being accessed more frequently than others throughout the fine-tuning process. Based on this insight, we introduce Vela, a novel fine-tuning system for MoE architectures that leverages expert locality to reduce communication overhead. Specifically, Vela implements a novel training and communication framework that separates expert layers from the MoE model, and employs a locality-aware expert placement mechanism to minimize the communication overhead, thereby significantly improving the fine-tuning efficiency. Our extensive array of evaluations demonstrates that Vela decreases the communication overhead by up to 25%, consequently accelerating the fine-tuning process by up to 28% compared to conventional methods.

Federated learning has garnered substantial research attention as a privacy-preserving learning paradigm. Nonetheless, its inherently distributed architecture, especially in asynchronous settings, poses vulnerabilities to model poisoning attacks. In such scenarios, malicious clients compromise the integrity of the global model by transmitting manipulated updates. To tackle this issue, previous research efforts like AFLGuard and Zeno++ have made strides but often hinge on the impractical assumption that the server has access to a clean dataset. In this study, we delve into practical solutions tailored for real-world scenarios, minimizing assumptions about the server's capabilities and accommodating varying settings. Specifically, we propose AsyncFilter module, which defends against poisoning attacks in asynchronous federated learning. Functioning as a plug-and-play module on the server, AsyncFilter enhances the learning process by statistically identifying and filtering out poisoned updates during training. On four real-world datasets, AsyncFilter effectively enhances global model accuracy against model poisoning attacks by up to 7%, 20%, 16% and 39% respectively. Through extensive evaluation, AsyncFilter demonstrates robust capabilities in detecting and mitigating model poisoning attacks in various scenarios encompassing diverse data and system heterogeneity, as well as varying attacker presence.

Federated learning has garnered significant research attention as a privacy-preserving learning paradigm. Asynchronous federated learning has been proposed to improve scalability by accommodating slower clients, commonly referred to as stragglers. However, asynchronous federated learning suffers from slow convergence with respect to wall-clock time, due to the existence of data heterogeneity and staleness. Existing strategies struggled to tackle both difficulties for a wide range of deep learning models. To address the problem, we proposePolaris, a theoretically sound design and a new take to client selection for asynchronous federated learning. WithPolaris, we first theoretically investigated the design space of client sampling strategies from a geometric optimization perspective, taking both data heterogeneity and staleness into account. Our design is not only theoretically proven, but also thoroughly tested in our reproducible experimental open-source testbed. Our experimental results demonstrates overwhelming evidence thatPolarisoutperformed existing state-of-the-art client selection strategies by a substantial margin over a wide variety of tasks and datasets, as we train image classification models using CIFAR-10, CIFAR-100, CINIC-10, Federated EMNIST, and a language modeling model using the Tiny Shakespeare dataset. Further, our extensive array of ablation studies have also shown thatPolarisis both scalable and robust as the size of datasets scale up and data heterogeneity vary.

Federated learning preserves data privacy by training machine learning models in a distributed fashion, where local models are trained on the client devices and aggregated on the server. Prevalent aggregation algorithms in federated learning perform well in homogeneous settings, but suffer from inadequate convergence in heterogeneous settings due to non-IID data distribution. In this paper, we explore the shortcomings of existing work and recognize that the memory loss of optimizers in aggregation steps limits convergence performance. In response, we propose FedRL, a new adaptive aggregation algorithm with the supervision of a policy-based deep reinforcement learning agent. Using real-world datasets, we evaluate the effectiveness of FedRL by comparing to state-of-the-art adaptive aggregation algorithms in the literature, and show its superiority in accelerating convergence to a target accuracy.