RHUL CS Cake Talks 🎂

Welcome to the RHUL CS Cake Talks webpage. Cake talks are informal seminars for postgraduate Computer Science students to discuss their research interests.

Seminars occur about once a month; the exact location and time may vary, so please check the schedule below. The format is a 30-minute talk, followed by a ~25-minute discussion session, followed by cake!

Each seminar is tagged with a general Topic and with Cake if cake is to be expected.

Click on Show Details to see the location, time and abstract of a seminar.

Time Speaker Title
Summer 2026
06 May 2026, 15:00 Sophie Hawkes  (Royal Holloway, University of London) Are Perceptual Hashes Safe to Share? Black-Box Inversion Attacks and Outsourced Private Set Intersection Cryptography Cake

Venue

McCrea 1-16

Short Bio

Sophie is a PhD researcher at Royal Holloway, University of London, within the CDT in Cyber Security for the Everyday, supervised by Dr Christian Weinert and Dr Maryam Mehrnezhad. Her research focuses on real-world applications of cryptography and privacy-preserving technologies, particularly outsourced private set intersection and the security and privacy of perceptual hashing in tools for reporting image-based abuse.

Abstract

Many digital systems for reporting sensitive content, such as non-consensually distributed intimate images (NCII) and other image-based sexual abuse (IBSA), rely on perceptual hashes as a robust digital fingerprint for detecting and removing images, without having to share the sensitive material itself. Our work challenges operational assumptions about the privacy of perceptual hashes, by demonstrating black-box hash inversion attacks that use Generative Adversarial Networks (GANs) to reconstruct approximate preimages from their perceptual hash values. We show these attacks are feasible with off-the-shelf technology against the most widely-deployed algorithms – PDQ (Facebook), NeuralHash (Apple), and PhotoDNA (Microsoft). Since these attacks exploit the availability of perceptual hash values in plaintext, the second half of the talk will discuss (outsourced) private set intersection (PSI) as a potential mitigation, to retain the functionality of secure hash matching whilst adding an extra layer of protection to prevent the exposure of vulnerable hash. Drawing from our recent Systematisation of Knowledge (SoK) paper, I will introduce a framework of 12 properties to help navigate trade-offs, security models, and design considerations within the somewhat fragmented landscape of outsourced PSI, in which resource-intensive computations (and sometimes dataset storage) are delegated to the cloud, and discuss research gaps and further work.

Spring 2026
01 Apr 2026, 15:00 Stephen Cook  (Royal Holloway, University of London) Security, Privacy, and Accessibility in Assistive IoT Technologies for Vulnerable Groups Usable Security and Privacy Cake

Venue

McCrea 1-16

Short Bio

Stephen Cook is a PhD researcher in the Information Security Group at Royal Holloway, University of London, supervised by Dr Maryam Mehrnezhad. His research sits at the intersection of security, privacy, and accessibility in health and assistive IoT technologies, with a particular focus on vulnerable user groups such as people with visual impairments. His completed work includes a security and privacy analysis of commercially available assistive IoT devices, accepted with major revisions to ACM Transactions on Accessible Computing, and an ongoing Systematisation of Knowledge on development practices in the Health IoT industry.

Abstract

Health Internet of Things (HIoT) technologies including wearable biosensors, smart mobility aids, and connected medical devices are increasingly relied upon by individuals with disabilities and long-term health conditions. However, these technologies introduce significant and often overlooked risks across three interconnected dimensions: security, privacy, and accessibility. This research investigates the interplay between these three factors within HIoT ecosystems, with particular focus on users with visual impairments. Through a pilot study of four commercially available assistive IoT devices, widespread vulnerabilities were identified, including unencrypted Bluetooth traffic, insecure authentication, excessive app permissions, and pervasive accessibility non-compliance. A complementary Systematisation of Knowledge (SoK) reveals that these shortcomings are driven by systemic development challenges — including limited regulatory awareness among developers, fragmented testing environments, and hardware constraints. For the final phase of this work we plan to conduct a large-scale study on the accessibility of iOS apps, as it remains one of the most widely used OS platforms by the VI community that has yet to be explored.

18 Mar 2026, 16:00 Georgii Potapov  (Royal Holloway, University of London) E-variables and tests of randomness with respect to families of distributions Algorithmic Randomness Cake

Venue

McCrea 1-16

Short Bio

Georgii is a PhD student at Royal Holloway working with Yuri Kalnishkan and Vladimir Vovk. His research interests range across Learning theory, Information theory, Computability, and Computational Complexity.

Abstract

E-variables are a relatively new approach for testing statistical hypotheses that has been experiencing major development during the last several years. E-variables were originally based on a concept of Levin's (average-bounded) randomness tests studied in Algorithmic Information Theory. We introduce the method of e-variable-approximability which can be used to develop a general approximation technique, allowing us to construct e-variables for popular distribution classes important for applications. We are going to discuss both this approach and how it helps us to reconnect theory of e-variables with Algorithmic Information Theory.

11 Feb 2026, 15:00 Cameron Jones  (Royal Holloway, University of London) Backwards Symbolic Execution Program Analysis Cake

Venue

McCrea 1-16

Short Bio

Cameron is a CDT/PhD student working towards Automated Vulnerability Verification/Proof of Concept generation. His focus is on whole tree Symbolic Execution (SE), specifically on methods to deal with the state space explosion which prevent SE from reaching far into real world programs. Currently he is attempting to compare the effectiveness of Backwards SE against targeted forwards SE.

Abstract

Forwards Symbolic Execution (SE) is a program analysis technique used in program verification, test case generation and vulnerability discovery/verification. SE functions by "executing" a program, similarly to an Interpreter, and following all possible paths of the program but storing all variables as a symbolic representation of their values, instead of the concrete (evaluated) values. These symbolic representations of all possible paths can then be used to validate the existence (or non-existence) of conditions in the program, generate test cases for complete branch coverage or find paths with desired properties. Duplication of the state to explore every path in a program becomes expensive fast and prevents analysis deep into programs, this is called a state space explosion. State space explosion is even more of an issue when there is a known target to be verified, since many of these states are irrelevant to the analysis but cannot always be identified as so. Backwards SE functions by starting at the target and performing SE backwards, providing the same output as forwards, however all states are relevant, just not always possible. This talk will briefly introduce SE, show how Backwards SE is different and then show some of the advantages and challenges it has compared to Forwards.

Winter 2025
03 Dec 2025, 14:00 Hayden Amarr  (Royal Holloway, University of London) Computing Optimal Labellings for Temporal Reachability Algorithms Cake

Venue

McCrea 0-02

Short Bio

Hayden is a first-year PhD student with research interests in temporal graph optimisation and game-theoretic problems. He is currently working on modelling networks that evolve over time, with the goal of improving efficiency and resource allocation in domains such as routing, scheduling, and optimisation-driven allocation.

Abstract

Most networks we think about are static, but real-world systems that are applicable to transportation logistics, communication, or social networks change dynamically over time. Research in temporal graphs aims to study how to model and analyse these dynamic systems. Whilst extensive research has already been conducted on static graphs, much remains to be understood about temporal graphs. Introducing a time dimension significantly alters many properties of static graphs, making it computationally more complex and intensive to perform operations that are easy on static graphs. This presentation will show how we can design valid schedules that meet the complex constraints on a temporal network. It will also explain why this task is computationally infeasible for general graphs with reasonable resources, and how this infeasibility motivates solving the problem for tractable graph structures like trees. A contribution of this research also includes a proof that bounds on the number of times each connection must be active to maintain complete, optimal reachability across the entire graph, where these graphs are restricted to trees. Finally, we will discuss the natural next step: moving from trees to graphs containing a single cycle and exploring how the complexity of the problem is impacted by the inclusion of cycles.

05 Nov 2025, 14:00 Philip Bogaars  (Royal Holloway, University of London) Network Restoration Games with Quotas Algorithms Cake

Venue

McCrea 0-02

Short Bio

Philip is a first-year PhD student researching humanitarian algorithms — computational methods designed to improve fairness, efficiency, and resource allocation in social and crisis settings.

Abstract

In a game of Network Restoration With Quotas, there is an underlying graph where a subset of its edges have to be restored by a set of agents. Each agent has a creation cost for each such edge, a traversal cost for every edge of the graph, and in addition they have a quota on the number of edges they have to restore. Then, given a set of edges that fulfill the quota, the cost of an agent is the cost of creating these edges, plus the cost of reaching them, i.e., the traversal cost. We prove that any cost-minimizing allocation is swap-stable, i.e., there is no profitable exchange of edges between any pair of agents, but computing one is hard even on trees. We complement this by designing an algorithm that finds a swap-stable allocation on trees in polynomial time and we quantify its cost against the optimal one.

Website source code