In here I will write down a few ideas that one can build their
MPhil/Part III on. The list is not containing all of the things I work
on a should be thought off as more of a guideline. If in doubt, just
send me an email and I am certain we can think of a project suiting
Student project ideas:
We have shown in “Hearing your
touch“ that attacks on virtual
keyboards are possible. Yet, the cross-device evaluation was
very limited. The project involves building a test bed similar
to the one described in the paper and collecting a dataset to be
later used for research.
Building effective acoustic defences to stop “Hearing your
touch“ from working.
assumed on the phone microphones for operation, yet, it might be
possible to perform the attacks with off-the-phone mics. The
project will involve building proof-of-concept of off-the-phone
done with an additional app installed on the phone. Yet, it
should be possible to perform similar attacks over the on-going
phone call. The project will involve analysis of what
preprocessing is required to stop attack from happening.
Last year we built a first version of BatNet —
ultrasound-based communication interface. The project involves
evaluating performance of different encoding schemes, that are
Signal-blocking sleeves became recently popular. The project
involves evaluating if signal-blocking sleeves are “air-gapped
enough” and there is actually no way to pass information around.
Prof Jeff Yan has introduced a technique called Differential
Imaging Forensics, it is
unclear thought how much information one can get out of it. The
project will involve evaluating the per-colour entropy of
shadows and evaluating the performance of the technique against
multi-frame shots found e.g. on youtube and iphone life photos.
It might be possible to do complete behind-the-camera scene
reconstruction based on the imperceptible shadows.
Recently, different agencies started employing ads monsters
like facebook and google to target particular types of people.
The project involves reverse engineering the types of people
It is known that companies like embedding tracking
information. Yet, the extend of how much it is done is unclear.
The project involves evaluating exactly that.
Spy cameras are becoming easier to buy and use. Those cameras
are really hard to humans to see, yet, there are certain
properties they have, that make them stand out for commonly used
smartphones. The project involves evaluating if commonly used
smartphones can be used to detect those spy cameras.
Attacking ML is not easy in the Black-box setting and
requires extensive oracle approximation. The project involves
developing mechanisms to stop attack transferability thought
disruption to approximation.
White-box based strong adversaries build the attacks using
multiple steps. Yet, those steps can be detected by the model
operator. The project involves building fingerprints of
different multi-step attacks. And showing how many steps one
need to detect that the attack is taking place.
We have shown in Taboo
Trap that we are capable of
detection adversarial samples with little effort. The detection
is based on the assumption that the data distribution of the
input matches the one of the evaluator. The project involves
empirically checking if in the real world it is indeed the case
and techniques like Taboo Trap or MagNet can actually be used.
We have shown in "Blackbox Attacks on Reinforcement Learning
Agents Using Approximated Temporal Information" that black-box
attacks are not easy. The project involves building RL-directed