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.

In regions of the world where censorship of the Internet is used to limit access to information, monitor the activity of Internet users, and quash dissent, anti-censorship proxies, or bridges, can offer a connection to the open Internet beyond a censor's area of influence. Bridge distribution systems, built to publicly distribute large pools of bridges to users in censored regions, face the inherent conflict of providing bridges to unknown users when some of them may be malicious. If not designed with care, bridge distribution systems can be quickly overwhelmed by attacks from censors, undermining the integrity of the system and the safety of users. It is therefore crucial to prioritize protecting users when developing such systems. In this paper, we present a new bridge distribution system, Lox. Lox prioritizes protecting the privacy of users and their social graphs and incorporates enumeration resistance mechanisms to improve access to bridges and limit the malicious behaviour of censors. We use an updated unlinkable multi-show anonymous credential scheme, suitable for a single credential issuer and verifier, to protect Lox bridge users and their social networks from being identified by malicious actors. We formalize a trust level scheme that is compatible with anonymous credentials and effectively limits malicious behaviour while maintaining user anonymity. Our work includes an open-sourced, Rust implementation of our Lox protocols as well as an evaluation of their performance. With reasonable performance and latency for the expected user base of our system, we demonstrate Lox as a practical, social graph protective bridge distribution system.

The salmon algorithm is a metaheuristic inspired by the behaviour of salmon swimming upstream to spawn. It has previously shown success when used for the creation of sets of robust tags for DNA sequencing applications, as well as for the travelling salesman problem. In this paper the salmon algorithm is evaluated for the construction of optimal covering and error-correcting codes, which are related to sequencing applications, as well as for the DNA fragment assembly problem, which is related to the travelling salesman problem. Parameter tuning for the salmon algorithm is extensively studied, as well as the use of automated parameter tuning.