While we value the benefits AI solutions provide, the downsides of it sometimes are forgotten or ignored. Society needs to understand better how effective different AI solutions are. The AI model robustness, possibility to manipulate the model make the AI solution not trusted. As well the ethical issues on its application make the AI application not effective. In most of the cases it can lead to some harms, rather than benefits for the AI solution clients or owners. this challenge invites creative persons to go deeper and work of estimation of the selected AI models robustness and its application ethical considerations.
The direction for the challenge ideas, which could be converted into multiple final thesis or individual project topics is the following:
- Backdoor attacks against Large Language Models.
- Jailbreaking attacks against Large Language Models.
- Detecting biases in LLM using adversarial prompts,
Regarding more details on the challenge and topics contact Antanas Čenys, VILNIUS TECH.