Next Seminar on 6.10.2022
Written on 06.10.2022 09:40 by Philip Lukert
The next seminar takes place on 12.10. at 14:30 (only Session A)
Session A: (14:30 - 15:00 && 15:00 - 15:30)
Erfan Balazadeh, Tejumade Afonja
Meeting-ID: 967 8620 5841
Session B does not exist next week
Speaker: Erfan Balazadeh
Type of talk: Bachelor Final
Advisor: Dr. Lucjan Hanzlik
Title: Timed-Release Cryptography using a Proof-of-Stake Blockchain
Research Area: 1
Abstract: Imagine a scenario where you want to encrypt a message, but you don't want it to be able to be decrypted by the receiving party right away.
The concept of "encrypting a message to the future" is not new and has been around for many years. The proposed solutions so far, like time-lock puzzles or verifiable delay functions, for instance,
are not perfect however. They require a lot of computing power and the speed can vary drastically depending on the hardware being used.
The thesis' goal was to implement a new encryption scheme, which is efficiently computable and which gets rid of the previously mentioned solutions' weaknesses, inside of a real-world setting.
The idea is to make use of the existing Proof-of-Stake architecture in the Ethereum 2.0 consensus protocol, where so called committees vote on new blocks by using an aggregatable signature scheme named BLS. One of the implementation tasks of the thesis was to see if it is possible to listen to the unaggregated BLS signatures and the signed message, which are necessary for the encryption scheme. Once enough of these unaggregated signatures are accumulated, we can go on to decrypt the message. Basically, a receiving party can only decrypt the message once certain conditions are met that the encrypter knows will happen in a desired amount of time in the future.
This talk will present the results and the findings of the thesis.
Speaker: Tejumade Afonja
Type of talk: Master Final
Advisor: Prof. Dr. Mario Fritz
Title: Learning Generative Models for Tabular Data based on Small Data
Research Area: Trustworthy Information Processing
Recent advances in generative modeling for images, speech, and natural language processing have also led to much interest in generative modeling for tabular data. However, tabular datasets are inherently heterogeneous and contain a mixture of numerical and categorical attributes, making them difficult to model. The current state-of-the-art tabular data generators (TDGs) have demonstrated impressive capabilities in capturing the statistical characteristics of the data, showing promising results in a few downstream machine learning tasks. However, existing results are based on large number of training instances (e.g., in tens of thousands), and are given only for specific metric, which rule out a myriad of practical scenarios where the sample size is limited and general properties beyond the specific metric are of interest. Hence, in this work, we systematically assessed the TDGs across various metrics as well as different subset sizes to better understand how these models behave in practical scenarios, specifically, the low-resource setting. To achieve this, we employ numerous existing measures that cover different aspects for the evaluation and propose two new metrics: the histogram intersection to measure the overlap between the synthetic and real data column, and the likelihood approximation to measure how likely the real data comes from the synthetic data distribution. Finally, we propose a benchmarking framework, faketrics, to comprehensively evaluate the TDGs along four axes: Utility, Joint, Column Pair, and Marginal so as to benchmark the evaluation of the models in low-resource setting.