Linear algebra is a cornerstone of neural network inference. The efficiency of popular frameworks, such as TensorFlow and PyTorch, critically depends on backend libraries providing highly optimized matrix multiplications and convolutions. A diverse range of these backends exists across platforms, including Intel MKL, Nvidia CUDA, and Apple Accelerate. Although these backends provide equivalent functionality, subtle variations in their implementations can lead to seemingly negligible differences during inference. In this paper, we investigate these minor discrepancies and demonstrate how they can be selectively amplified by adversaries. Specifically, we introduce *Chimera examples*, inputs to models that elicit conflicting predictions depending on the employed backend library. These inputs can even be constructed with integer values, creating a vulnerability exploitable from real-world input domains. We analyze the prevalence and extent of the underlying attack surface and propose corresponding defenses to mitigate this threat.

Compression is an effective technique for fitting data in available memory, reducing I/O, and increasing instruction parallelism. While data systems primarily rely on lossless compression, modern machine learning (ML) systems exploit the approximate nature of ML and mostly use lossy compression via low-precision floating- or fixed-point representations. The resulting unknown impact on learning progress, and model accuracy, however, create trust concerns, that require trial and error, and are problematic for declarative ML pipelines. Given the trend towards increasingly complex, composite ML pipelines---with outer loops for hyper-parameter tuning, feature selection, and data cleaning/augmentation---it is hard for a user to infer the impact of lossy compression. Sparsity exploitation is a common lossless scheme used to improve performance without this uncertainty. Evolving this concept to general redundancy-exploiting compression is a natural next step. Existing work on lossless compression and compressed linear algebra (CLA) enable such exploitation to a degree, but face challenges for general applicability. In this paper, we address these limitations with a workload-aware compression framework, comprising a broad spectrum of new compression schemes and kernels. Instead of a data-centric approach that optimizes compression ratios, our workload-aware compression summarizes the workload of an ML pipeline, and optimizes the compression and execution plan to minimize execution time. On various micro benchmarks and end-to-end ML pipelines, we observe improvements for individual operations up to 10,000x and ML algorithms up to νmprint6.6 x compared to uncompressed operations.

Federated learning allows training machine learning (ML) models without central consolidation of the raw data. Variants of such federated learning systems enable privacy-preserving ML, and address data ownership and/or sharing constraints. However, existing work mostly adopt data-parallel parameter-server architectures for mini-batch training, require manual construction of federated runtime plans, and largely ignore the broad variety of data preparation, ML algorithms, and model debugging. Over the last years, we extended Apache SystemDS by an additional federated runtime backend for federated linear-algebra programs, federated parameter servers, and federated data preparation. In this paper, we share the system-level compiler and runtime integration, new features such as multi-tenant federated learning, selected federated primitives, multi-key homomorphic encryption, and our monitoring infrastructure. Our demonstrator showcases how composite ML pipelines can be compiled into federated runtime plans with low overhead.

Data science workflows are largely exploratory, dealing with under-specified objectives, open-ended problems, and unknown business value. Therefore, little investment is made in systematic acquisition, integration, and pre-processing of data. This lack of infrastructure results in redundant manual effort and computation. Furthermore, central data consolidation is not always technically or economically desirable or even feasible (e.g., due to privacy, and/or data ownership). The ExDRa system aims to provide system infrastructure for this exploratory data science process on federated and heterogeneous, raw data sources. Technical focus areas include (1) ad-hoc and federated data integration on raw data, (2) data organization and reuse of intermediates, and (3) optimization of the data science lifecycle, under awareness of partially accessible data. In this paper, we describe use cases, the overall system architecture, selected features of SystemDS' new federated backend (for federated linear algebra programs, federated parameter servers, and federated data preparation), as well as promising initial results. Beyond existing work on federated learning, ExDRa focuses on enterprise federated ML and related data pre-processing challenges. In this context, federated ML has the potential to create a more fine-grained spectrum of data ownership and thus, even new markets.