Opportunity brief
Identify Synthetic Media, Train Robust ML Models, and Safeguard Digital Integrity!
About the Event
Dive into the cutting edge of AI ethics and security in this high-intensity machine learning hackathon focused on Deepfake Detection!
Participants will compete to build accurate, efficient, and robust models designed to identify synthetic media. Designed to test both technical prowess and ethical AI thinking, this single-round challenge demands originality and precision as you train and optimize models against a provided custom dataset.
Submission Guidelines
Submissions must be provided via a Google Drive folder link (set to "Anyone with the link can view") containing:
- Source Code: Zipped code directory or a public GitHub repository link with a comprehensive README.md. - Trained Model Weights: Model files compatible with PyTorch or TensorFlow. - Prediction CSV: Model outputs generated on the official test set. - Detailed Documentation: Covering your approach, model architecture, hyperparameter selection, training pipeline, and challenges faced. - Explanatory Video (Optional): A brief 3–5 minute video demonstrating your solution and methodology.
Rules & Regulations
- Dataset Restrictions: Only the officially provided dataset may be used for training. External training datasets are strictly prohibited. - Pre-trained Models: Pre-trained architectures are permitted, but fine-tuning must be executed exclusively on the provided dataset. - Reproducibility: Submitted solutions must be fully reproducible. Incomplete submissions or unverified code will not be evaluated. - Originality: All work must be original. Code plagiarism or theft will result in immediate disqualification. - Final Submission: Only one final submission per team is evaluated (teams may update their files until the deadline, but only the latest upload counts).
Evaluation Criteria
- Accuracy on Test Set (50%): Performance and accuracy metrics on the unlabelled test data. - Model Efficiency (20%): Inference speed and memory footprint during evaluation. - Generalization & Robustness (20%): Model stability and resistance to edge-case noise/artifacts. - Clarity of Documentation (10%): Quality, readability, and depth of the submitted technical report.