Defence of doctoral thesis in the field of Computer Science, FM Tommi Gröndahl

Natural language processing (NLP) is among the most widely used types of artificial
intelligence. Among its many applications, it has been adopted for detecting adversarial
material such as deception or hate speech. However, text classification also
brings about possibilities for censorship and violations of privacy. In this doctoral dissertation
I compare leading NLP methods from the perspective of security and privacy
in particular. I focus on text classification, automatic text generation, and different
ways of representing semantic content. I also propose technical improvements to
prior methods.

The results indicate that text classification is typically based on rather simple features,
such as the presence of certain words. Complex machine learning models get
caught on such features as well as simpler models. This makes them vulnerable to
attacks that attempt to confuse the classifier without significantly changing the text’s
content. On the other hand, such evasion of classification provides a defence against
e.g. author identification, which could have major consequences for the author’s privacy
and safety.

From a technical perspective, I compare deep neural networks (DNNs) to more traditional
rule-based methods. Major differences are found, and their benefits and challenges
are commonly opposite. Even though DNNs can process language data far
more broadly than manually programmed rules, they have trouble maintaining
semantic content in a detailed way. Therefore DNNs alone are not always the optimal
solution, and combining the best aspects of different methods is crucial.

Opponent: professor Jörg Tiedemann, University of Helsinki, Finland

Custos: professor N. Asokan, University of Waterloo / Aalto University School of Science, Department of Computer Science

Doctoral candidate’s contact information: [email protected]

The defence will be organised via remote technology (Zoom). Link to the event. 

Zoom Quick Guide

The doctoral thesis will be publicly displayed 10 days before the defence in the Aaltodoc publication archive of Aalto University.

Electronic thesis. (aaltodoc.aalto.fi)