Phantom Transfer: Data Poisoning can Survive Data-Level Defences

1. 1Build a dataset with subliminal learning modifications to demonstrate the Phantom Transfer attack
2. 2Run experiments to test the attack's effectiveness against different models and targets
3. 3Configure data-level defenses to attempt to filter out the poison
4. 4Test the robustness of the attack against these defenses
5. 5Apply the insights from the attack to improve model security and robustness

AI engineers and data scientists on a team benefit from understanding this attack to improve model security and robustness, as it can inform the development of more effective defenses

💡 Phantom Transfer attack can modify subliminal learning to work in real-world contexts, making it difficult to filter out even with knowledge of the poison

🚨 New data poisoning attack: Phantom Transfer can survive data-level defenses! 🤖