Training machine learning models on privacy-sensitive data has become a popular practice, driving innovation in ever-expanding fields. This has opened the door to new attacks that can have serious privacy implications. One such attack, the Membership Inference Attack (MIA), exposes whether or not a particular data point was used to train a model. A growing body of literature uses Differentially Private (DP) training algorithms as a defence against such attacks. However, these works evaluate the defence under the restrictive assumption that all members of the training set, as well as non-members, are independent and identically distributed. This assumption does not hold for many real-world use cases in the literature. Motivated by this, we evaluate membership inference with statistical dependencies among samples and explain why DP does not provide meaningful protection (the privacy parameter$\epsilon$scales with the training set size$n$) in this more general case. We conduct a series of empirical evaluations with off-the-shelf MIAs using training sets built from real-world data showing different types of dependencies among samples. Our results reveal that training set dependencies can severely increase the performance of MIAs, and therefore assuming that data samples are statistically independent can significantly underestimate the performance of MIAs.

Differential privacy (DP) allows data analysts to query databases that contain users' sensitive information while providing a quantifiable privacy guarantee to users. Recent interactive DP systems such as APEx provide accuracy guarantees over the query responses, but fail to support a large number of queries with a limited total privacy budget, as they process incoming queries independently from past queries. We present an interactive, accuracy-aware DP query engine, CacheDP , which utilizes a differentially private cache of past responses, to answer the current workload at a lower privacy budget, while meeting strict accuracy guarantees. We integrate complex DP mechanisms with our structured cache, through novel cache-aware DP cost optimization. Our thorough evaluation illustrates that CacheDP can accurately answer various workload sequences, while lowering the privacy loss as compared to related work.

Devices like smartphones and laptops employ some form of user authentication to ensure that access to confidential data by the wrong user is avoided. Implicit authentication aims to limit the number of explicit authentications that a user is subjected to by using passive approaches to authenticate the user. Context detection frameworks aim to reduce explicit authentications by disabling explicit authentication entirely when appropriate. Since explicit and implicit authentication are not mutually exclusive, we can also use context detection frameworks to decide whether explicit or implicit authentication should be used when authentication is required. We present a novel context detection framework, PUPy, that uses sensed context data to infer and make available three values–privacy, unfamiliarity, and proximity–allowing clients of our framework, like authentication services, to better adapt to different contexts. As opposed to existing work, our context detection framework is based on an optimistic approach to context detection. Our assumption is that the absence of data, like the inability to detect nearby people or devices, can be taken as a sign that a context is safe. Such an optimistic approach may provide less security than a pessimistic approach, but provides a significantly improved user experience due to reducing the number of explicit authentications. We provide an Android implementation of the framework, including an API that allows other developers to contribute modules to the system. We also conduct a statistical analysis of our framework based on a large real-world dataset. We find that PUPy compares favourably to existing works, permitting a 77.2% reduction in the number of explicit authentications.