Federated Learning (FL) enables the training of machine learning models across decentralized clients holding local data samples without exchanging them. While this paradigm preserves data privacy, it is vulnerable to adversarial attacks from malicious clients that can poison the global model by submitting harmful updates. Label-flipping attacks, where attackers intentionally mislabel their training data, pose a significant threat due to their subtlety and similarity to legitimate updates. In this work, we introduce a lightweight and effective filtering system that leverages distance and similarity metrics between client updates and a trimmed mean to identify and exclude malicious updates. Through experiments on the CIFAR-10 dataset with non-independent and identically distributed (Non-IID) data distributions, we demonstrate that our approach improves the robustness of the global model against label-flipping attacks, maintaining accuracy levels close to those achieved without any attackers.

The tendency for cell fate to be robust to most perturbations, yet sensitive to certain perturbations raises intriguing questions about the existence of a key path within the underlying molecular network that critically determines distinct cell fates. Reprogramming and trans-differentiation clearly show examples of cell fate change by regulating only a few or even a single molecular switch. However, it is still unknown how to identify such a switch, called a master regulator, and how cell fate is determined by its regulation. Here, we present CAESAR, a computational framework that can systematically identify master regulators and unravel the resulting canalizing kernel, a key substructure of interconnected feedbacks that is critical for cell fate determination. We demonstrate that CAESAR can successfully predict reprogramming factors for de-differentiation into mouse embryonic stem cells and trans-differentiation of hematopoietic stem cells, while unveiling the underlying essential mechanism through the canalizing kernel. CAESAR provides a system-level understanding of how complex molecular networks determine cell fates.

Providing low latency to end-users while maximizing server utilization and system throughput is crucial for cloud ML servers. NVIDIA's recently announced Ampere GPU architecture provides features to "reconfigure" one large, monolithic GPU into multiple smaller "GPU partitions". Such feature provides cloud ML service providers the ability to utilize the reconfigurable GPU not only for large-batch training but also for small-batch inference with the potential to achieve high resource utilization. We study this emerging GPU architecture with reconfigurability to develop a high-performance multi-GPU ML inference server, presenting a sophisticated partitioning algorithm for reconfigurable GPUs combined with an elastic scheduling algorithm tailored for our heterogeneously partitioned GPU server.

In cloud ML inference systems, batching is an essential technique to increase throughput which helps optimize total-cost-of-ownership. Prior graph batching combines the individual DNN graphs into a single one, allowing multiple inputs to be concurrently executed in parallel. We observe that the coarse-grained graph batching becomes suboptimal in effectively handling the dynamic inference request traffic, leaving significant performance left on the table. This paper proposes LazyBatching, an SLA-aware batching system that considers both scheduling and batching in the granularity of individual graph nodes, rather than the entire graph for flexible batching. We show that LazyBatching can intelligently determine the set of nodes that can be efficiently batched together, achieving an average 15×, 1.5×, and 5.5 × improvement than graph batching in terms of average response time, throughput, and SLA satisfaction, respectively.