Research Focus: Week of February 6, 2023

Published

Microsoft Research Focus 09 edition, week of February 6, 2023

Welcome to Research Focus, a new series of blog posts that highlights notable publications, events, code/datasets, new hires and other milestones from across the research community at Microsoft.

Behind the Tech podcast with Tobi Lütke: CEO and Founder, Shopify

In the latest episode of Behind the Tech, Microsoft CTO Kevin Scott is joined by Tobi Lütke, CEO and founder of the Canadian multinational e-commerce platform Shopify. Since his early days running an online snowboard shop from his carport, Tobi has envisioned himself as a craftsman first and a business exec second, a mindset he has used to solve a wide variety of problems. He and Kevin discuss applying computer science and engineering techniques to build and scale a company, the idea of bringing an ‘apprentice mindset’ to his work, and how Tobi’s daily practice of writing code and tinkering in his home lab inspires him to be a more creative leader.

Tune in now to enjoy the discussion.


Distribution inference risks: Identifying and mitigating sources of leakage

Distribution inference (or property inference) attacks allow an adversary to infer distributional information about the training data of a machine learning model, which can cause significant problems. For example, leaking distribution of sensitive attributes such as gender or race can create a serious privacy concern. This kind of attack has been shown to be feasible on different types of models and datasets. However, little attention has been given to identifying the potential causes of such leakages and to proposing mitigations.

A new paper, Distribution Inference Risks: Identifying and Mitigating Sources of Leakage, focuses on theoretically and empirically analyzing the sources of information leakage that allow an adversary to perpetrate distribution inference attacks. The researchers identified three sources of leakage: (1) memorizing specific information about the value of interest to the adversary; (2) wrong inductive bias of the model; and (3) finiteness of the training data. Next, based on their analysis, the researchers propose principled mitigation techniques against distribution inference attacks. Specifically, they demonstrate that causal learning techniques are more resilient to a particular type of distribution inference risk — distributional membership inference — than associative learning methods. And lastly, they present a formalization of distribution inference that allows for reasoning about more general adversaries than was previously possible.


Siva Kakarla wins Applied Networking Research Prize

Microsoft’s Siva Kakarla (opens in new tab) has been awarded an Applied Networking Research Prize (opens in new tab) for 2023 in recognition of his work on checking the correctness of nameservers. A senior researcher in the Networking Research Group (opens in new tab) of Microsoft Research, Kakarla was one of six people to receive this annual award from the Internet Research Task Force (opens in new tab).

In their paper: SCALE: Automatically Finding RFC Compliance Bugs in DNS Nameservers, Kakarla and his colleagues introduce the first approach for finding RFC (request for comment) compliance errors in DNS nameserver implementations through automatic test generation. Their approach, called Small-scope Constraint-driven Automated Logical Execution, or SCALE, generates high-coverage tests for covering RFC behaviors.

The Applied Networking Research Prize acknowledges advances in applied networking, interesting new research ideas of potential relevance to the internet standards community, and people that are likely to have an impact on internet standards and technologies.

Related publications

Continue reading

See all blog posts