The Event#
The Singapore AI CTF was a jeopardy-style AI/ML Capture-the-Flag (CTF) competition organised by GovTech’s Cybersecurity Group (CSG) and Dreadnode. In this competition, participants tackle AI security challenges across 7 domains to collect ‘flags’ in the form of hidden files or information.
The event’s page on CTFTime: https://ctftime.org/event/2546
Challenge Categories#
The challenges spanned the following 7 categories:
Data Analysis#
Participants need to utilise various data analysis and visualisation tools to extract hidden information within a given dataset.
Model Extraction#
Participants use malicious means to extract information from the model (e.g. parameters, architecture, or training data)
Model Fingerprinting#
Participants must verify watermarks and/or fingerprints in AI models to verify their authenticity and detect unauthorised copies.
Prompt Injection#
An attack where hackers use a specially crafted input to trick AI into disclosing a hidden secret.
Adversarial AI Image#
An attacker injects malicious data into the training set causing AI to misclassify an image.
Adversarial Audio Generation#
Involves replicating voices and audio using AI.
Model Inversion#
An attacker aims to infer/reconstruct input data based on model outputs.