Secure AI Hackathon Day 2

The Secure AI Hackathon Day 2 challenges teams to safeguard and explore safety dynamics within AI models. The scenario places participants in the role of a consortium of five banks that need to detect fraud and network intrusions without pooling customer data due to legal restrictions. The solution involves federated learning (FL), where each bank trains a model locally on its own data, sharing only trained model updates rather than raw data to form a global model.

This event focuses on two specific tracks:

• Intermediate Track: Addresses the "non-IID" problem where different banks see different kinds of traffic, causing naive federated learning to fail. Participants must design a better aggregation strategy to close the performance gap caused by this data skew.
• Advanced Track (Optional): Explores adversarial settings where one of the "banks" turns malicious and sends sabotaged updates. Teams must either detect this poisoning attack or make their system resistant to it.

Unlike Day 1, which used an automated leaderboard, these tracks are evaluated through a combination of objective metrics and qualitative assessment. Submissions require a Kaggle Writeup (max 1,500 words) detailing the approach, results, and F1 score comparisons (before vs. after). Additionally, teams must provide a Media Gallery with a cover image, a public Notebook with all cells executed successfully, a public video (3 minutes or less) acting as an elevator pitch, and a public GitHub repository containing the final exported model and metrics file for independent verification.

Judging is conducted by CAIRLab mentors based on four criteria totaling 100 points: Automated Detection Score (40%), Writeup Quality (30%), Live Demo & Presentation (20%), and Creativity & Security Mindset (10%). Awards include "The Gallant Giant" for the Intermediate track and "The Wholesome Expert" for the Advanced track. Top teams may receive an invitation to ongoing CAIRLab research.