Modern machine learning systems use models trained on ever-growing corpora.
Typically, metadata such as ownership, access control, or licensing information
is ignored during training. Instead, to mitigate privacy risks, we rely on
generic techniques such as dataset sanitization and differentially private
model training, with inherent privacy/utility trade-offs that hurt model
performance. Moreover, these techniques have limitations in scenarios where
sensitive information is shared across multiple participants and fine-grained
access control is required. By ignoring metadata, we therefore miss an
opportunity to better address security, privacy, and confidentiality
challenges. In this paper, we take an information flow control perspective to
describe machine learning systems, which allows us to leverage metadata such as
access control policies and define clear-cut privacy and confidentiality
guarantees with interpretable information flows. Under this perspective, we
contrast two different approaches to achieve user-level non-interference: 1)
fine-tuning per-user models, and 2) retrieval augmented models that access
user-specific datasets at inference time. We compare these two approaches to a
trivially non-interfering zero-shot baseline using a public model and to a
baseline that fine-tunes this model on the whole corpus. We evaluate trained
models on two datasets of scientific articles and demonstrate that retrieval
augmented architectures deliver the best utility, scalability, and flexibility
while satisfying strict non-interference guarantees.

在本文中，我们从信息流控制的角度描述了机器学习系统，利用元数据，如访问控制策略来定义明确的隐私和机密性保证，通过比较两种不同方法，即针对每个用户进行微调模型和在推理时访问用户特定数据集的检索增强模型，我们证明检索增强架构可以在满足严格的非干扰性保证的同时提供最佳的效用、可扩展性和灵活性。

从信息流控制的角度重新思考机器学习管道中的隐私

Rethinking Privacy in Machine Learning Pipelines from an Information  Flow Control Perspective

In today's machine learning (ML) models, any part of the training data can
affect its output. This lack of control for information flow from training data
to model output is a major obstacle in training models on sensitive data when
access control only allows individual users to access a subset of data. To
enable secure machine learning for access controlled data, we propose the
notion of information flow control for machine learning, and develop a secure
Transformer-based language model based on the Mixture-of-Experts (MoE)
architecture. The secure MoE architecture controls information flow by limiting
the influence of training data from each security domain to a single expert
module, and only enabling a subset of experts at inference time based on an
access control policy. The evaluation using a large corpus of text data shows
that the proposed MoE architecture has minimal (1.9%) performance overhead and
can significantly improve model accuracy (up to 37%) by enabling training on
access-controlled data.

为了使受访者能够更加安全地使用 Access-controlled 数据，该研究提出了信息流控制的概念并开发了一种基于 Mixture-of-Experts 的 Transformer 语言模型。实验结果表明，该模型能够使数据的性能开销最小化并显著提高训练模型的准确性。

模块化模型架构中的机器学习信息流控制

Information Flow Control in Machine Learning through Modular Model  Architecture

A model of cloud services is emerging whereby a few trusted providers manage
the underlying hardware and communications whereas many companies build on this
infrastructure to offer higher level, cloud-hosted PaaS services and/or SaaS
applications. From the start, strong isolation between cloud tenants was seen
to be of paramount importance, provided first by virtual machines (VM) and
later by containers, which share the operating system (OS) kernel. Increasingly
it is the case that applications also require facilities to effect isolation
and protection of data managed by those applications. They also require
flexible data sharing with other applications, often across the traditional
cloud-isolation boundaries; for example, when government provides many related
services for its citizens on a common platform. Similar considerations apply to
the end-users of applications. But in particular, the incorporation of cloud
services within `Internet of Things' architectures is driving the requirements
for both protection and cross-application data sharing.
These concerns relate to the management of data. Traditional access control
is application and principal/role specific, applied at policy enforcement
points, after which there is no subsequent control over where data flows; a
crucial issue once data has left its owner's control by cloud-hosted
applications and within cloud-services. Information Flow Control (IFC), in
addition, offers system-wide, end-to-end, flow control based on the properties
of the data. We discuss the potential of cloud-deployed IFC for enforcing
owners' dataflow policy with regard to protection and sharing, as well as
safeguarding against malicious or buggy software. In addition, the audit log
associated with IFC provides transparency, giving configurable system-wide
visibility over data flows. [...]

本研究探讨了云数据安全的问题，提出了利用信息流控制 (IFC) 实现数据保护与共享的方法，为云服务及物联网架构提供了新思路。